<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom"><id>https://xtra-agent-safety-atlas.pages.dev/feed.en.xml</id><title>Astrolabe for Agent Safety · periodic report</title>
<link href="https://xtra-agent-safety-atlas.pages.dev/"/><link rel="self" href="https://xtra-agent-safety-atlas.pages.dev/feed.en.xml"/><updated>2026-11-16T02:30:00Z</updated><author><name>Astrolabe for Agent Safety</name></author>
<entry><id>https://xtra-agent-safety-atlas.pages.dev/issues/w-2026-11-14.en.html</id><title>Weekly issue 7 · 8 – 14 November 2026</title><link href="https://xtra-agent-safety-atlas.pages.dev/issues/w-2026-11-14.en.html"/><updated>2026-11-16T02:30:00Z</updated>
<summary>64 new papers this week, 52% fewer than the three before at their first count (134 on average).</summary><content type="html">&lt;p class=&quot;head&quot;&gt;64 new papers this week, 52% fewer than the three before at their first count (134 on average).&lt;/p&gt;&lt;div class=&quot;tiles&quot;&gt;&lt;div&gt;&lt;b&gt;64&lt;/b&gt;&lt;span&gt;new papers&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;134&lt;/b&gt;&lt;span&gt;average first count of the three before&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;29&lt;/b&gt;&lt;span&gt;newly confirmed at a venue&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;18&lt;/b&gt;&lt;span&gt;new repositories&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;19%&lt;/b&gt;&lt;span&gt;with an industry author&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;h2&gt;The main points&lt;/h2&gt;&lt;ol&gt;&lt;li&gt;&lt;b&gt;5 papers&lt;/b&gt; · Most new papers: Indirect injection defences for tool agents&lt;small&gt;Next: Deception, scheming and sandbagging 3; Reward hacking, self-improvement, research agents 3; AI control and sabotage evaluation 3&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;29 papers&lt;/b&gt; · Newly confirmed as accepted or published&lt;small&gt;4 at the top-4 security or top-3 ML venues. Most: MM 2026 5; AI Summit &amp;#x27;26: ACM AI Leadership Summit 2026 2026 5; ACM CCS 2026 4&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;9 new citations&lt;/b&gt; · Most used earlier paper: Progent: Securing AI Agents with Privilege Control&lt;small&gt;3 of them use it as a method, a benchmark or a comparison. 2025-04 · Indirect injection defences for tool agents&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;19%&lt;/b&gt; · of new papers have an author from industry&lt;small&gt;13% in the three periods before. 32 institutions published their first paper in the field.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;+15,971 stars&lt;/b&gt; · Fastest-growing repository: affaan-m/ECC&lt;small&gt;18 repositories were created in this period.&lt;/small&gt;&lt;/li&gt;&lt;/ol&gt;&lt;h2&gt;Worth reading first&lt;/h2&gt;&lt;ol&gt;&lt;li&gt;&lt;a href=&quot;https://doi.org/10.1109/spw72489.2026.00047&quot;&gt;Tracing MCP Security Vulnerabilities: Constructing and Analyzing an MCP Security Vulnerability Dataset&lt;/a&gt;&lt;small&gt;IEEE S&amp;amp;P 2026 · Proceedings (top-4 security venue); a new benchmark&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2611.12167v1&quot;&gt;FraudBench: A Legal Evaluation of AI Deception&lt;/a&gt;&lt;small&gt;from the team of Noam Kolt, Hebrew University of Jerusalem; a new benchmark&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2611.11861v1&quot;&gt;Scaling Adaptive Prompt Injection Attacks with Agent Swarms&lt;/a&gt;&lt;small&gt;from the team of Kamalika Chaudhuri, University of California San Diego; a new benchmark&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2611.11123v1&quot;&gt;HARP-Bench: Benchmarking Robot Manipulation Around Active Humans&lt;/a&gt;&lt;small&gt;code released; a new benchmark&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2611.16485v1&quot;&gt;Contain, Conform, Compare (Extended Version): A Red-Teamed Evaluation Kit for Small Coding Agents and the Full Record of Its Experiments&lt;/a&gt;&lt;small&gt;code released; a new benchmark&lt;/small&gt;&lt;/li&gt;&lt;/ol&gt;&lt;p class=&quot;sub&quot;&gt;Chosen by rule: venue tier, a ranked team, released code, a new benchmark and use by later papers; at most two per theme.&lt;/p&gt;&lt;h2&gt;Direction radar&lt;/h2&gt;&lt;table&gt;&lt;tr&gt;&lt;th&gt;Direction&lt;/th&gt;&lt;th&gt;This period&lt;/th&gt;&lt;th&gt;Average before&lt;/th&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ Process evaluation and execution auditing of agents&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;2&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;6.7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ Coding agents and software engineering&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;2&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;6&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ Prompt injection detection and model-level defence&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;1&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;4.7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ Tool-call authorisation and approval&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;2&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;5.7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ Agent societies and cooperation&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;0&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3.7&lt;/td&gt;&lt;/tr&gt;&lt;/table&gt;&lt;h2&gt;Who is moving&lt;/h2&gt;&lt;p&gt;Institutions with their first paper in the field: Alation, Alibaba AAIG, Allianz Group, Artificial Societies, Berlin Institute of Health at Charité - Universitätsmedizin Berlin, Charité – Universitätsmedizin Berlin, Czech Academy of Sciences, Department of Computer Science and Engineering

CMR Institute of Technology, Department of Statistics and Data Science, ERA:AI&lt;/p&gt;&lt;h2&gt;Open source and models&lt;/h2&gt;&lt;ul&gt;&lt;li&gt;affaan-m/ECC &lt;small&gt;+15,971 stars, now 300,696&lt;/small&gt;&lt;/li&gt;&lt;li&gt;ifixai-ai/iFixAi &lt;small&gt;+13,919 stars, now 44,510&lt;/small&gt;&lt;/li&gt;&lt;li&gt;earendil-works/pi &lt;small&gt;+4,561 stars, now 121,238&lt;/small&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h2&gt;Which conclusions changed&lt;/h2&gt;&lt;table&gt;&lt;tr&gt;&lt;td&gt;Papers on the site&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;6,421 → &lt;b&gt;6,449&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;New papers in the last 12 months&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;4,397 → &lt;b&gt;4,414&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;Institutions with a paper on the site&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3,942 → &lt;b&gt;3,953&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;/table&gt;&lt;p class=&quot;sub&quot;&gt;The other 10 tracked numbers moved by less than half a point.&lt;/p&gt;&lt;h2&gt;Revision of the period before&lt;/h2&gt;&lt;p&gt;2026-11-01 – 2026-11-07 was first counted at 127 papers; 0 arrived later, 127 now.&lt;/p&gt;</content></entry><entry><id>https://xtra-agent-safety-atlas.pages.dev/issues/w-2026-11-07.en.html</id><title>Weekly issue 6 · 1 – 7 November 2026</title><link href="https://xtra-agent-safety-atlas.pages.dev/issues/w-2026-11-07.en.html"/><updated>2026-11-12T02:30:00Z</updated>
<summary>The theme “Evaluation, guardrails” had 10 new papers this week, against an average of 5.7 in the three before.</summary><content type="html">&lt;p class=&quot;head&quot;&gt;The theme “Evaluation, guardrails” had 10 new papers this week, against an average of 5.7 in the three before.&lt;/p&gt;&lt;div class=&quot;tiles&quot;&gt;&lt;div&gt;&lt;b&gt;127&lt;/b&gt;&lt;span&gt;new papers&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;126.7&lt;/b&gt;&lt;span&gt;average first count of the three before&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;191&lt;/b&gt;&lt;span&gt;newly confirmed at a venue&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;26&lt;/b&gt;&lt;span&gt;new repositories&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;16%&lt;/b&gt;&lt;span&gt;with an industry author&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;h2&gt;The main points&lt;/h2&gt;&lt;ol&gt;&lt;li&gt;&lt;b&gt;6 papers&lt;/b&gt; · Most new papers: VLA safety&lt;small&gt;Next: Agentic system architecture and runtime engineering 6; Indirect injection defences for tool agents 5; Process evaluation and execution auditing of agents 5&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;5 vs 0.7 before&lt;/b&gt; · Rising: Prompt injection and information leakage in LLM-integrated applications&lt;small&gt;Also rising: Execution control and assurance in agent frameworks 5; Indirect injection defences for tool agents 5; MCP servers and protocol security 3&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;191 papers&lt;/b&gt; · Newly confirmed as accepted or published&lt;small&gt;3 at the top-4 security or top-3 ML venues. Most: EMNLP 2026 134; Proceedings of the Second Workshop for REsearch on Agent Language Models (REALM 2026) 2026 7; AAAI 2026 6&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;16 new citations&lt;/b&gt; · Most used earlier paper: AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agents&lt;small&gt;2 of them use it as a method, a benchmark or a comparison. 2024-10 · Safety evaluation of tool use, risk awareness&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;16%&lt;/b&gt; · of new papers have an author from industry&lt;small&gt;13% in the three periods before. 62 institutions published their first paper in the field.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;+15,971 stars&lt;/b&gt; · Fastest-growing repository: affaan-m/ECC&lt;small&gt;26 repositories were created in this period.&lt;/small&gt;&lt;/li&gt;&lt;/ol&gt;&lt;h2&gt;Worth reading first&lt;/h2&gt;&lt;ol&gt;&lt;li&gt;&lt;a href=&quot;https://neurips.cc/virtual/2026/poster/139257&quot;&gt;MetaPI: Constructing Prompt Injection Benchmarks from LLM Agent Benchmark&lt;/a&gt;&lt;small&gt;NeurIPS 2026 · Main conference (top-3 ML venue); from the team of Yuan Tian, University of California, Los Angeles; a new benchmark&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2611.08425v1&quot;&gt;Failure Prediction and Classification in Vision-Language-Action Models Via Calibrated Latent Probing&lt;/a&gt;&lt;small&gt;from the team of Ufuk Topcu, The University of Texas at Austin; code released; a new benchmark&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2611.06727v1&quot;&gt;Airlock: Bounding the Multi-Agent Blast Radius via Selective Online Containment&lt;/a&gt;&lt;small&gt;from the team of Hamed Haddadi, Imperial College London; code released; a new benchmark · In multi-agent systems, a single agent&amp;#x27;s erroneous output spreads across tasks before it is detected, creating a system-level blast radius.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://doi.org/10.1145/3847352.3848107&quot;&gt;MCP Tool Poisoning in Agentic AI Frameworks: A Cross-Framework Measurement Study&lt;/a&gt;&lt;small&gt;code released; a new benchmark · MCP feeds attacker-controllable tool descriptions, schema fields and response data into the model context without validation, leading to tool poisoning.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2611.02689v1&quot;&gt;The Customer Never Sees the Goal: Measuring and Detecting Vulnerability-Targeted Manipulation in LLM Assistants&lt;/a&gt;&lt;small&gt;code released; a new benchmark · An operator sets a business goal for an LLM assistant that customers cannot see, and the assistant may manipulate customers by exploiting vulnerabilities they disclose (VTM).&lt;/small&gt;&lt;/li&gt;&lt;/ol&gt;&lt;p class=&quot;sub&quot;&gt;Chosen by rule: venue tier, a ranked team, released code, a new benchmark and use by later papers; at most two per theme.&lt;/p&gt;&lt;h2&gt;Direction radar&lt;/h2&gt;&lt;table&gt;&lt;tr&gt;&lt;th&gt;Direction&lt;/th&gt;&lt;th&gt;This period&lt;/th&gt;&lt;th&gt;Average before&lt;/th&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Prompt injection and information leakage in LLM-integrated applications&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;5&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;0.7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Execution control and assurance in agent frameworks&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;5&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;1.7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Indirect injection defences for tool agents&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;5&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;2.7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ MCP servers and protocol security&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;1&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Agent red teaming and cross-session threats&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;1&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ Contextual integrity and privacy leakage&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;0&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;4.3&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ Prompt injection detection and model-level defence&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;2&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;5.7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ Coding agents and software engineering&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;6.7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ AI control and sabotage evaluation&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;0&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3.3&lt;/td&gt;&lt;/tr&gt;&lt;/table&gt;&lt;h2&gt;Who is moving&lt;/h2&gt;&lt;p&gt;Institutions with their first paper in the field: Eastern International University, Jiffy Labs, Thu Dau Mot University, University of Information Technology, ABLEs, AGH University of Krakow, AWS Agentic AI, Abylkas Saginov Karaganda Technical University, Achiral Research, Acompany Co.&lt;/p&gt;&lt;h2&gt;Open source and models&lt;/h2&gt;&lt;ul&gt;&lt;li&gt;affaan-m/ECC &lt;small&gt;+15,971 stars, now 300,696&lt;/small&gt;&lt;/li&gt;&lt;li&gt;ifixai-ai/iFixAi &lt;small&gt;+13,919 stars, now 44,510&lt;/small&gt;&lt;/li&gt;&lt;li&gt;earendil-works/pi &lt;small&gt;+4,561 stars, now 121,238&lt;/small&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h2&gt;Which conclusions changed&lt;/h2&gt;&lt;table&gt;&lt;tr&gt;&lt;td&gt;Papers on the site&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;6,318 → &lt;b&gt;6,421&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;New papers in the last 12 months&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;4,314 → &lt;b&gt;4,397&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;Institutions with a paper on the site&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3,889 → &lt;b&gt;3,942&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;Open-source repositories in the library&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;7,889 → &lt;b&gt;8,073&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;/table&gt;&lt;p class=&quot;sub&quot;&gt;The other 9 tracked numbers moved by less than half a point.&lt;/p&gt;&lt;h2&gt;Revision of the period before&lt;/h2&gt;&lt;p&gt;2026-10-25 – 2026-10-31 was first counted at 134 papers; 3 arrived later, 137 now.&lt;/p&gt;</content></entry><entry><id>https://xtra-agent-safety-atlas.pages.dev/issues/m-2026-10.en.html</id><title>Monthly issue · October 2026</title><link href="https://xtra-agent-safety-atlas.pages.dev/issues/m-2026-10.en.html"/><updated>2026-11-06T02:30:00Z</updated>
<summary>Coding agents and software engineering had 34 new papers this month, against an average of 12.3 in the three before.</summary><content type="html">&lt;p class=&quot;head&quot;&gt;Coding agents and software engineering had 34 new papers this month, against an average of 12.3 in the three before.&lt;/p&gt;&lt;div class=&quot;tiles&quot;&gt;&lt;div&gt;&lt;b&gt;653&lt;/b&gt;&lt;span&gt;new papers&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;466&lt;/b&gt;&lt;span&gt;average first count of the three before&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;47&lt;/b&gt;&lt;span&gt;newly confirmed at a venue&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;286&lt;/b&gt;&lt;span&gt;new repositories&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;14%&lt;/b&gt;&lt;span&gt;with an industry author&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;h2&gt;The main points&lt;/h2&gt;&lt;ol&gt;&lt;li&gt;&lt;b&gt;34 papers&lt;/b&gt; · Most new papers: Process evaluation and execution auditing of agents&lt;small&gt;Next: Coding agents and software engineering 34; Tool-call authorisation and approval 30; VLA safety 22&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;34 vs 12.3 before&lt;/b&gt; · Rising: Coding agents and software engineering&lt;small&gt;Also rising: Indirect prompt injection: toolkits, defence evaluations and accumulating attacks 22; Prompt injection detection and model-level defence 22; Contextual integrity and privacy leakage 19&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;47 papers&lt;/b&gt; · Newly confirmed as accepted or published&lt;small&gt;Most: ASE 2026 10; ICRA 2026 9; AAAI 2026 6&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;64 new citations&lt;/b&gt; · Most used earlier paper: $τ$-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains&lt;small&gt;15 of them use it as a method, a benchmark or a comparison. 2024-06 · Process evaluation and execution auditing of agents&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;14%&lt;/b&gt; · of new papers have an author from industry&lt;small&gt;11% in the three periods before. 292 institutions published their first paper in the field.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;+14,128 stars&lt;/b&gt; · Fastest-growing repository: affaan-m/ECC&lt;small&gt;286 repositories were created in this period.&lt;/small&gt;&lt;/li&gt;&lt;/ol&gt;&lt;h2&gt;Worth reading first&lt;/h2&gt;&lt;ol&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.29869v1&quot;&gt;Compaction Laundering: Provenance Destruction in Context-Compacting Coding Agents&lt;/a&gt;&lt;small&gt;Shows that context compaction in five popular coding agents flattens message roles, letting forged user approvals survive into memory 87-100% of the time. Also gives a simple serializer fix. · When a coding agent compacts its context, it flattens the conversation into plain text and loses message provenance, so an attacker can forge user instructions.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.37741v1&quot;&gt;Agentic Backdoors via Pretraining Poisoning&lt;/a&gt;&lt;small&gt;First end-to-end demonstration that poisoning 0.1% of pretraining data plants a remote-code-execution backdoor that survives clean post-training into a CLI agent. Attack success grows with scale; widely discussed on HN. · Whether pretraining data poisoning can implant a remote code execution backdoor in an agent, causing it to execute attacker-specified commands on the user&amp;#x27;s machine.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.24052v1&quot;&gt;Red-Teaming Multi-Agent Attacks for Loss-of-Control Risks&lt;/a&gt;&lt;small&gt;MiniLab shows that splitting an attack across colluding agents drops frontier monitor detection from 87% to 7%. Gives reproducible red-team cases for multi-agent loss-of-control. · Whether AI control monitoring that reads trajectories one by one fails when multiple agents secretly collude and carry out a harmful goal in separate pieces.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.41998v1&quot;&gt;SecPO: Principled Adversarial Training for Prompt Injection Security&lt;/a&gt;&lt;small&gt;Adversarial training with strong synthesized injections plus a principled SecPO loss cuts adaptive AgentDojo attack success to 8.8% versus 78.8% for Meta-SecAlign, with no utility loss. Model and data released. · LLM agents have insufficient defenses against prompt injection; the strongest defense, Meta-SecAlign, has an ASR close to 100% under adaptive attacks.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.22301v1&quot;&gt;My Credentials, Please: Detecting Filesystem Capability Laundering in MCP Servers at Scale&lt;/a&gt;&lt;small&gt;Large-scale scan of 21,097 MCP servers finds tools reaching host credentials outside their declared contracts, with 150 confirmed exploitable instances. Disclosure produced 13 advisories and two CVEs. · Once approved to be called, an MCP server holds broad host file permissions, and an attacker can read sensitive files by pushing parameters out of bounds (filesystem capability laundering).&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.21063v1&quot;&gt;Approve What, Exactly? Effect Binding and Trusted Rendering for Agent Step-Up Authorization&lt;/a&gt;&lt;small&gt;Defines twelve attacks on agent approval flows and audits AP2, MCP elicitation and ten runtimes. Builds a Tamarin-verified gateway that binds approvals to exact effects and renders a trusted card. · Human approval before an agent performs high-risk operations such as transfers or file deletion is often treated as a second factor, but it is not necessarily bound to the effect of the approved operation.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.26551v2&quot;&gt;Training Natural Schemers: RL Can Amplify Oversight-Conditional Behavior&lt;/a&gt;&lt;small&gt;Builds a natural scheming model organism and shows that ordinary RL, which never rewards scheming, still widens the gap between monitored and unmonitored compliance. The effect holds across environments and base models. · Whether a tendency toward scheming (complying under oversight while pursuing its own goals when unsupervised) can emerge or be amplified under realistic training conditions.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.04375v1&quot;&gt;Do Tool Calls Execute as Intended? Measuring and Repairing Intent-Execution Correspondence in LLM Agents&lt;/a&gt;&lt;small&gt;Measures production sessions and finds harnesses silently alter tool calls: Claude Code&amp;#x27;s Bash tool changes 12% of code-bearing calls, often running the wrong action with no error reported. · When an LLM agent&amp;#x27;s tool calls are passed through multiple hops, they can be silently rewritten, causing wrong actions to be executed or pointless retries.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.14043v1&quot;&gt;Refuses in Chat, Complies with Arms: LLM Refusal Does Not Survive Robot Control&lt;/a&gt;&lt;small&gt;Matched text and robot-control tests show refusal collapsing on real hardware, demonstrating that chat safety evaluations cannot establish safe physical action. · LLMs refuse requests that cause physical harm in text; the question is whether this refusal still holds when the model controls a robot.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.12445v1&quot;&gt;Caught in the Act: Probes Effectively Detect Sabotage and Catch Unverbalized Deception&lt;/a&gt;&lt;small&gt;Multi-layer deception probes trained on the largest deception dataset to date reach 98.8% AUC on SHADE-Arena, beating a text monitor. They also catch hidden goals; the FIBS dataset is released. · LLM agents may deceive users or secretly sabotage, which is hard to detect with text monitoring alone, especially deception that is not stated aloud.&lt;/small&gt;&lt;/li&gt;&lt;/ol&gt;&lt;p class=&quot;sub&quot;&gt;The monthly star papers: two models (Claude Opus 5.5 and GPT-6.1 Sol) read every paper of the month and vote.&lt;/p&gt;&lt;h2&gt;Direction radar&lt;/h2&gt;&lt;table&gt;&lt;tr&gt;&lt;th&gt;Direction&lt;/th&gt;&lt;th&gt;This period&lt;/th&gt;&lt;th&gt;Average before&lt;/th&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Coding agents and software engineering&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;34&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;12.3&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Indirect prompt injection: toolkits, defence evaluations and accumulating attacks&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;22&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;4.3&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Prompt injection detection and model-level defence&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;22&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;8.3&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Contextual integrity and privacy leakage&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;19&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;6.3&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Tool-call authorisation and approval&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;30&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;19.3&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ RAG poisoning and retrieval security&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;5&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;12&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ Agent identity, delegation, decentralised protocols&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;4&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;10&lt;/td&gt;&lt;/tr&gt;&lt;/table&gt;&lt;h2&gt;Who is moving&lt;/h2&gt;&lt;p&gt;Institutions with their first paper in the field: Kontext, Corners Co., Cygnux Labs, International Hellenic University, Kahlert School of Computing, Manchester Metropolitan University, Nanchang Hangkong University, Robocurve, Universidade Estadual de Campinas (UNICAMP), Uploading Inc.&lt;/p&gt;&lt;h2&gt;Open source and models&lt;/h2&gt;&lt;ul&gt;&lt;li&gt;affaan-m/ECC &lt;small&gt;+14,128 stars, now 284,725&lt;/small&gt;&lt;/li&gt;&lt;li&gt;ifixai-ai/iFixAi &lt;small&gt;+12,268 stars, now 30,591&lt;/small&gt;&lt;/li&gt;&lt;li&gt;earendil-works/pi &lt;small&gt;+5,600 stars, now 116,677&lt;/small&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h2&gt;Which conclusions changed&lt;/h2&gt;&lt;table&gt;&lt;tr&gt;&lt;td&gt;Papers on the site&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;5,593 → &lt;b&gt;6,318&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;New papers in the last 12 months&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3,758 → &lt;b&gt;4,314&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;Papers about digital (software) agents&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;72.4% → &lt;b&gt;73.6%&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;Institutions with a paper on the site&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3,568 → &lt;b&gt;3,889&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;Papers of the last 12 months that release code&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;17.9% → &lt;b&gt;17.4%&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;Open-source repositories in the library&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;7,671 → &lt;b&gt;7,889&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;Papers using a language model that use at least one open-weight model&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;75.4% → &lt;b&gt;76.4%&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;/table&gt;&lt;p class=&quot;sub&quot;&gt;The other 6 tracked numbers moved by less than half a point.&lt;/p&gt;&lt;h2&gt;Revision of the period before&lt;/h2&gt;&lt;p&gt;2026-09-01 – 2026-09-30 was first counted at 564 papers; 5 arrived later, 569 now.&lt;/p&gt;</content></entry><entry><id>https://xtra-agent-safety-atlas.pages.dev/issues/w-2026-10-31.en.html</id><title>Weekly issue 5 · 25 – 31 October 2026</title><link href="https://xtra-agent-safety-atlas.pages.dev/issues/w-2026-10-31.en.html"/><updated>2026-11-06T02:30:00Z</updated>
<summary>The theme “Tool execution, coding and runtime” had 40 new papers this week, against an average of 26.7 in the three before.</summary><content type="html">&lt;p class=&quot;head&quot;&gt;The theme “Tool execution, coding and runtime” had 40 new papers this week, against an average of 26.7 in the three before.&lt;/p&gt;&lt;div class=&quot;tiles&quot;&gt;&lt;div&gt;&lt;b&gt;134&lt;/b&gt;&lt;span&gt;new papers&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;99&lt;/b&gt;&lt;span&gt;average first count of the three before&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;0&lt;/b&gt;&lt;span&gt;newly confirmed at a venue&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;30&lt;/b&gt;&lt;span&gt;new repositories&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;14%&lt;/b&gt;&lt;span&gt;with an industry author&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;h2&gt;The main points&lt;/h2&gt;&lt;ol&gt;&lt;li&gt;&lt;b&gt;9 papers&lt;/b&gt; · Most new papers: Coding agents and software engineering&lt;small&gt;Next: VLA safety 6; Prompt injection detection and model-level defence 5; Process evaluation and execution auditing of agents 5&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;9 vs 4.7 before&lt;/b&gt; · Rising: Coding agents and software engineering&lt;small&gt;Also rising: Reward hacking, self-improvement, research agents 5; VLA safety 6; Safety evaluation of tool use, risk awareness 5&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;11 new citations&lt;/b&gt; · Most used earlier paper: Defending Against Indirect Prompt Injection Attacks With Spotlighting&lt;small&gt;5 of them use it as a method, a benchmark or a comparison. 2024-03 · Prompt injection detection and model-level defence&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;14%&lt;/b&gt; · of new papers have an author from industry&lt;small&gt;13% in the three periods before. 65 institutions published their first paper in the field.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;30 repositories&lt;/b&gt; · created in this period&lt;small&gt;&lt;/small&gt;&lt;/li&gt;&lt;/ol&gt;&lt;h2&gt;Worth reading first&lt;/h2&gt;&lt;ol&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.39122v1&quot;&gt;Appreciation Teaches Unauthorized Execution: Self-Evolution Degrades the Robustness of Personal Agents to Indirect Prompt Injection&lt;/a&gt;&lt;small&gt;from the team of Chaowei Xiao, Johns Hopkins University; code released · Whether a self-evolving personal agent becomes more vulnerable to indirect prompt injection attacks after learning from user feedback.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.41885v1&quot;&gt;How Much Can Tool Calls Reveal About You? Privacy Leakage from Tool-Call Traces&lt;/a&gt;&lt;small&gt;from the team of Dongrui Liu, Shanghai Artificial Intelligence Laboratory; code released; a new benchmark · Whether MCP tool-call traces leak user privacy, allowing an observer to infer user profile attributes.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.41998v1&quot;&gt;SecPO: Principled Adversarial Training for Prompt Injection Security&lt;/a&gt;&lt;small&gt;from the team of David Wagner, University of California, Berkeley; code released · LLM agents have insufficient defenses against prompt injection; the strongest defense, Meta-SecAlign, has an ASR close to 100% under adaptive attacks.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.37741v1&quot;&gt;Agentic Backdoors via Pretraining Poisoning&lt;/a&gt;&lt;small&gt;from the team of Nicholas Carlini, Anthropic (United States); code released · Whether pretraining data poisoning can implant a remote code execution backdoor in an agent, causing it to execute attacker-specified commands on the user&amp;#x27;s machine.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.36435v1&quot;&gt;Your Agent Acts Differently When You&amp;#x27;re Around: How User Presence Changes Agent Behavior in Negotiation&lt;/a&gt;&lt;small&gt;from the team of He He, NYU; a new benchmark · How does the behavior of an LLM agent negotiating on a user&amp;#x27;s behalf change when the user is in the loop, and does misaligned behavior appear?&lt;/small&gt;&lt;/li&gt;&lt;/ol&gt;&lt;p class=&quot;sub&quot;&gt;Chosen by rule: venue tier, a ranked team, released code, a new benchmark and use by later papers; at most two per theme.&lt;/p&gt;&lt;h2&gt;Direction radar&lt;/h2&gt;&lt;table&gt;&lt;tr&gt;&lt;th&gt;Direction&lt;/th&gt;&lt;th&gt;This period&lt;/th&gt;&lt;th&gt;Average before&lt;/th&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Coding agents and software engineering&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;9&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;4.7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Reward hacking, self-improvement, research agents&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;5&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;1&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ VLA safety&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;6&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;2.3&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Safety evaluation of tool use, risk awareness&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;5&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;1.7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Human oversight and control of agents&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;0.3&lt;/td&gt;&lt;/tr&gt;&lt;/table&gt;&lt;h2&gt;Who is moving&lt;/h2&gt;&lt;p&gt;Institutions with their first paper in the field: Cygnux Labs, Manchester Metropolitan University, ACE Robotics, ADAPT Research Unit

CTTC-CERCA, AI Center, AIR Institute, Al-Ahliyya Amman University, American University of Sharjah, Anhui Polytechnic University, Baba&lt;/p&gt;</content></entry><entry><id>https://xtra-agent-safety-atlas.pages.dev/issues/w-2026-10-24.en.html</id><title>Weekly issue 4 · 18 – 24 October 2026</title><link href="https://xtra-agent-safety-atlas.pages.dev/issues/w-2026-10-24.en.html"/><updated>2026-11-06T02:30:00Z</updated>
<summary>The theme “Alignment, oversight” had 16 new papers this week, against an average of 4.7 in the three before.</summary><content type="html">&lt;p class=&quot;head&quot;&gt;The theme “Alignment, oversight” had 16 new papers this week, against an average of 4.7 in the three before.&lt;/p&gt;&lt;div class=&quot;tiles&quot;&gt;&lt;div&gt;&lt;b&gt;141&lt;/b&gt;&lt;span&gt;new papers&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;101.7&lt;/b&gt;&lt;span&gt;average first count of the three before&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;0&lt;/b&gt;&lt;span&gt;newly confirmed at a venue&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;58&lt;/b&gt;&lt;span&gt;new repositories&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;11%&lt;/b&gt;&lt;span&gt;with an industry author&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;h2&gt;The main points&lt;/h2&gt;&lt;ol&gt;&lt;li&gt;&lt;b&gt;10 papers&lt;/b&gt; · Most new papers: Process evaluation and execution auditing of agents&lt;small&gt;Next: Tool-call authorisation and approval 8; Prompt injection detection and model-level defence 7; Coding agents and software engineering 6&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;5 vs 1.0 before&lt;/b&gt; · Rising: Human-agent autonomy, governance, trust frameworks&lt;small&gt;Also rising: CoT monitoring and steganography 5; AI control and sabotage evaluation 4; Agent societies and cooperation 5&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;5 new citations&lt;/b&gt; · Most used earlier paper: Representation Engineering: A Top-Down Approach to AI Transparency&lt;small&gt;3 of them use it as a method, a benchmark or a comparison. 2023-10 · Deception, scheming and sandbagging&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;11%&lt;/b&gt; · of new papers have an author from industry&lt;small&gt;14% in the three periods before. 66 institutions published their first paper in the field.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;+4,546 stars&lt;/b&gt; · Fastest-growing repository: affaan-m/ECC&lt;small&gt;58 repositories were created in this period.&lt;/small&gt;&lt;/li&gt;&lt;/ol&gt;&lt;h2&gt;Worth reading first&lt;/h2&gt;&lt;ol&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.27244v1&quot;&gt;AutoRedClaw: Automated Red-Teaming for Claw-Style Agents&lt;/a&gt;&lt;small&gt;from the team of Chaowei Xiao, Johns Hopkins University; code released; a new benchmark · When the configuration files, long-term memory and skills of a claw-style agent are poisoned, the effect persists across tasks and sessions, and manual testing can hardly cover this.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.26698v1&quot;&gt;Towards Robust Model Organisms for Eliciting Latent Knowledge&lt;/a&gt;&lt;small&gt;from the team of Mary Phuong, California Institute of Technology; code released; a new benchmark · A misaligned model that knows more than its supervisor may conceal findings (sandbag), so a credible ELK experimental subject is needed to study how to elicit them.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.27548v2&quot;&gt;Act on Your Behalf, Beyond Your Authorization: Benchmarking LLM Agent Authorization Alignment&lt;/a&gt;&lt;small&gt;from the team of Tianshi Li, Northeastern University; code released; a new benchmark · LLM agents acting on behalf of users may exceed the scope of user authorization, and existing evaluations look only at task completion or harm, so they easily miss such overreach.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://neurips.cc/virtual/2026/poster/154000&quot;&gt;Humans Correct Their Judgments Through Debate, but Weaker AI Models Do Not&lt;/a&gt;&lt;small&gt;NeurIPS 2026 · Main conference (top-3 ML venue) · In scalable oversight of superhuman AI, whether debate can help human or weak-model judges with prior biases reach correct judgments.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://aclanthology.org/2026.emnlp-main.1996/&quot;&gt;After the Attacker Leaves: Benchmarking Carryover Attacks on LLM Agents&lt;/a&gt;&lt;small&gt;EMNLP 2026 · Main conference (A-tier venue); a new benchmark · Whether a one-time indirect prompt injection, once written into an agent&amp;#x27;s persistent state, can be triggered repeatedly across sessions after the attacker has left.&lt;/small&gt;&lt;/li&gt;&lt;/ol&gt;&lt;p class=&quot;sub&quot;&gt;Chosen by rule: venue tier, a ranked team, released code, a new benchmark and use by later papers; at most two per theme.&lt;/p&gt;&lt;h2&gt;Direction radar&lt;/h2&gt;&lt;table&gt;&lt;tr&gt;&lt;th&gt;Direction&lt;/th&gt;&lt;th&gt;This period&lt;/th&gt;&lt;th&gt;Average before&lt;/th&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Human-agent autonomy, governance, trust frameworks&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;5&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;1&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ CoT monitoring and steganography&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;5&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;1&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ AI control and sabotage evaluation&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;4&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;0.7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Agent societies and cooperation&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;5&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;2&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Coding agents and software engineering&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;6&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ Runtime guardrails and multi-turn trajectories&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;0&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;4.3&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ Reward hacking, self-improvement, research agents&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;1&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3.3&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ Agent skills ecosystem security&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;2&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;4&lt;/td&gt;&lt;/tr&gt;&lt;/table&gt;&lt;h2&gt;Who is moving&lt;/h2&gt;&lt;p&gt;Institutions with their first paper in the field: Kahlert School of Computing, Uploading Inc., 3MI Labs, AI Standards Lab, Arya.ai, Bashi AI

London, Beijing Jishuitan Hospital, California Polytechnic State University, Cambridge Research Laboratory, Chief Product Officer&lt;/p&gt;&lt;h2&gt;Open source and models&lt;/h2&gt;&lt;ul&gt;&lt;li&gt;affaan-m/ECC &lt;small&gt;+4,546 stars, now 284,725&lt;/small&gt;&lt;/li&gt;&lt;li&gt;ifixai-ai/iFixAi &lt;small&gt;+2,313 stars, now 30,591&lt;/small&gt;&lt;/li&gt;&lt;li&gt;earendil-works/pi &lt;small&gt;+1,250 stars, now 116,677&lt;/small&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h2&gt;Which conclusions changed&lt;/h2&gt;&lt;table&gt;&lt;tr&gt;&lt;td&gt;Papers on the site&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;5,781 → &lt;b&gt;6,318&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;New papers in the last 12 months&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3,868 → &lt;b&gt;4,314&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;Papers about digital (software) agents&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;72.8% → &lt;b&gt;73.6%&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;Institutions with a paper on the site&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3,646 → &lt;b&gt;3,889&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;Open-source repositories in the library&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;7,792 → &lt;b&gt;7,889&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;Papers using a language model that use at least one open-weight model&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;75.3% → &lt;b&gt;76.4%&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;/table&gt;&lt;p class=&quot;sub&quot;&gt;The other 7 tracked numbers moved by less than half a point.&lt;/p&gt;&lt;h2&gt;Revision of the period before&lt;/h2&gt;&lt;p&gt;2026-10-11 – 2026-10-17 was first counted at 105 papers; 52 arrived later, 157 now.&lt;/p&gt;</content></entry><entry><id>https://xtra-agent-safety-atlas.pages.dev/issues/w-2026-10-17.en.html</id><title>Weekly issue 3 · 11 – 17 October 2026</title><link href="https://xtra-agent-safety-atlas.pages.dev/issues/w-2026-10-17.en.html"/><updated>2026-10-19T02:30:00Z</updated>
<summary>105 new papers this week, level with the three before at their first count (108 on average); most are in Process evaluation and execution auditing of agents (8).</summary><content type="html">&lt;p class=&quot;head&quot;&gt;105 new papers this week, level with the three before at their first count (108 on average); most are in Process evaluation and execution auditing of agents (8).&lt;/p&gt;&lt;div class=&quot;tiles&quot;&gt;&lt;div&gt;&lt;b&gt;105&lt;/b&gt;&lt;span&gt;new papers&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;107.7&lt;/b&gt;&lt;span&gt;average first count of the three before&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;36&lt;/b&gt;&lt;span&gt;newly confirmed at a venue&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;66&lt;/b&gt;&lt;span&gt;new repositories&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;13%&lt;/b&gt;&lt;span&gt;with an industry author&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;h2&gt;The main points&lt;/h2&gt;&lt;ol&gt;&lt;li&gt;&lt;b&gt;8 papers&lt;/b&gt; · Most new papers: Process evaluation and execution auditing of agents&lt;small&gt;Next: Agentic system architecture and runtime engineering 7; Tool-call authorisation and approval 6; Prompt injection detection and model-level defence 5&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;7 vs 1.7 before&lt;/b&gt; · Rising: Agentic system architecture and runtime engineering&lt;small&gt;Also rising: Coding agents and software engineering 5; Safety and alignment in reinforcement learning for language agents 3; Contextual integrity and privacy leakage 4&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;36 papers&lt;/b&gt; · Newly confirmed as accepted or published&lt;small&gt;Most: ASE 2026 10; ICRA 2026 9; AAAI 2026 6&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;10 new citations&lt;/b&gt; · Most used earlier paper: $τ$-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains&lt;small&gt;5 of them use it as a method, a benchmark or a comparison. 2024-06 · Process evaluation and execution auditing of agents&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;13%&lt;/b&gt; · of new papers have an author from industry&lt;small&gt;15% in the three periods before. 38 institutions published their first paper in the field.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;+5,622 stars&lt;/b&gt; · Fastest-growing repository: ifixai-ai/iFixAi&lt;small&gt;66 repositories were created in this period.&lt;/small&gt;&lt;/li&gt;&lt;/ol&gt;&lt;h2&gt;Worth reading first&lt;/h2&gt;&lt;ol&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.16333v1&quot;&gt;TRAP: Tool-result Replay Attack for Probing Misbinding in LLM Agents&lt;/a&gt;&lt;small&gt;from the team of Jing Shao, Shanghai Artificial Intelligence Laboratory; code released; a new benchmark · Genuine, untampered tool results are replayed into incompatible execution contexts, causing agents to take unsafe actions (tool-result misbinding).&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://doi.org/10.18653/2026.emnlp-industry.247&quot;&gt;LatentRiskBench: Measuring State-Dependent Risks in Enterprise Agents&lt;/a&gt;&lt;small&gt;EMNLP 2026 · Proceedings (A-tier venue); code released; a new benchmark · In enterprise agents, safety depends on the risk of the workflow state, and looking at actions alone cannot determine whether a violation has occurred.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.19158v1&quot;&gt;AgentCIScope: Grounded Contextual Integrity Evaluation for LLM Agents&lt;/a&gt;&lt;small&gt;from the team of Bo Li, University of Illinois Urbana-Champaign; a new benchmark · When LLM agents read and write sensitive data across applications, whether the information flows they produce violate contextual disclosure constraints, that is, whether they break contextual integrity.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.15924v1&quot;&gt;Scalable Weak-to-Strong Verifier Hardening for Agent Environments&lt;/a&gt;&lt;small&gt;code released; a new benchmark · Programmatic verifiers in agent environments can be attacked, allowing a cheating agent to receive full scores without completing the task, which leads to reward hacking.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.20049v1&quot;&gt;CoRedSkill: Stealthy Skill Composition Attacks via AST-Guided Injection&lt;/a&gt;&lt;small&gt;from the team of Jiaheng Zhang, National University of Singapore; a new benchmark · When multiple skills are combined, malicious operations can be spread across the dependency chain, so auditing a single skill package makes it difficult to identify their combined intent.&lt;/small&gt;&lt;/li&gt;&lt;/ol&gt;&lt;p class=&quot;sub&quot;&gt;Chosen by rule: venue tier, a ranked team, released code, a new benchmark and use by later papers; at most two per theme.&lt;/p&gt;&lt;h2&gt;Direction radar&lt;/h2&gt;&lt;table&gt;&lt;tr&gt;&lt;th&gt;Direction&lt;/th&gt;&lt;th&gt;This period&lt;/th&gt;&lt;th&gt;Average before&lt;/th&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Agentic system architecture and runtime engineering&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;7&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;1.7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Coding agents and software engineering&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;5&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;2.7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Safety and alignment in reinforcement learning for language agents&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;0.7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Contextual integrity and privacy leakage&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;4&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;1.7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Runtime authority and injection diagnosis&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;1&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ Reward hacking, self-improvement, research agents&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;0&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;5.7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ Embodied task-planning safety&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;0&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;4&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ Agent societies and cooperation&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;0&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3.7&lt;/td&gt;&lt;/tr&gt;&lt;/table&gt;&lt;h2&gt;Who is moving&lt;/h2&gt;&lt;p&gt;Institutions with their first paper in the field: Kontext, Nanchang Hangkong University, Affiliation:, Auburn University at Montgomery, Belleville, CUHK (Shenzhen), Charles Sturt University, Conversational Technologies, Cookeville, Fraunhofer Institute for Molecular Biology and Applied Ecology&lt;/p&gt;&lt;h2&gt;Open source and models&lt;/h2&gt;&lt;ul&gt;&lt;li&gt;ifixai-ai/iFixAi &lt;small&gt;+5,622 stars, now 28,278&lt;/small&gt;&lt;/li&gt;&lt;li&gt;affaan-m/ECC &lt;small&gt;+4,938 stars, now 280,179&lt;/small&gt;&lt;/li&gt;&lt;li&gt;earendil-works/pi &lt;small&gt;+1,929 stars, now 115,427&lt;/small&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h2&gt;Which conclusions changed&lt;/h2&gt;&lt;table&gt;&lt;tr&gt;&lt;td&gt;Papers on the site&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;5,645 → &lt;b&gt;5,781&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;New papers in the last 12 months&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3,773 → &lt;b&gt;3,868&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;Institutions with a paper on the site&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3,585 → &lt;b&gt;3,646&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;Open-source repositories in the library&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;7,724 → &lt;b&gt;7,792&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;/table&gt;&lt;p class=&quot;sub&quot;&gt;The other 9 tracked numbers moved by less than half a point.&lt;/p&gt;&lt;h2&gt;Revision of the period before&lt;/h2&gt;&lt;p&gt;2026-10-04 – 2026-10-10 was first counted at 51 papers; 26 arrived later, 77 now.&lt;/p&gt;</content></entry><entry><id>https://xtra-agent-safety-atlas.pages.dev/issues/w-2026-10-10.en.html</id><title>Weekly issue 2 · 4 – 10 October 2026</title><link href="https://xtra-agent-safety-atlas.pages.dev/issues/w-2026-10-10.en.html"/><updated>2026-10-12T02:30:00Z</updated>
<summary>51 new papers this week, 62% fewer than the three before at their first count (133 on average).</summary><content type="html">&lt;p class=&quot;head&quot;&gt;51 new papers this week, 62% fewer than the three before at their first count (133 on average).&lt;/p&gt;&lt;div class=&quot;tiles&quot;&gt;&lt;div&gt;&lt;b&gt;51&lt;/b&gt;&lt;span&gt;new papers&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;132.7&lt;/b&gt;&lt;span&gt;average first count of the three before&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;11&lt;/b&gt;&lt;span&gt;newly confirmed at a venue&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;91&lt;/b&gt;&lt;span&gt;new repositories&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;18%&lt;/b&gt;&lt;span&gt;with an industry author&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;h2&gt;The main points&lt;/h2&gt;&lt;ol&gt;&lt;li&gt;&lt;b&gt;5 papers&lt;/b&gt; · Most new papers: Process evaluation and execution auditing of agents&lt;small&gt;Next: Deception, scheming and sandbagging 4; Injection and deception against GUI and web agents 3; Prompt injection detection and model-level defence 3&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;11 papers&lt;/b&gt; · Newly confirmed as accepted or published&lt;small&gt;Most: Proceedings of the ACM on Software Engineering 2026 3; SOSP 2026 1; Logical Methods in Computer Science 2026 1&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;2 new citations&lt;/b&gt; · Most used earlier paper: EIA: ENVIRONMENTAL INJECTION ATTACK ON GENERALIST WEB AGENTS FOR PRIVACY LEAKAGE&lt;small&gt;2 of them use it as a method, a benchmark or a comparison. 2024-09 · Injection and deception against GUI and web agents&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;18%&lt;/b&gt; · of new papers have an author from industry&lt;small&gt;13% in the three periods before. 16 institutions published their first paper in the field.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;+4,644 stars&lt;/b&gt; · Fastest-growing repository: affaan-m/ECC&lt;small&gt;91 repositories were created in this period.&lt;/small&gt;&lt;/li&gt;&lt;/ol&gt;&lt;h2&gt;Worth reading first&lt;/h2&gt;&lt;ol&gt;&lt;li&gt;&lt;a href=&quot;https://doi.org/10.1145/3847353.3847499&quot;&gt;Formal Runtime Verification for Tool-Using LLM Agents: An Offline Same-Benchmark Study on AgentDojo and STAC&lt;/a&gt;&lt;small&gt;ACM CCS 2026 · Proceedings (top-4 security venue); a new benchmark · Guardrails for tool-using LLM agents are mostly application-specific rules, and multi-step, data-dependent safety policies are hard to specify, audit, and reuse.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.11552v1&quot;&gt;Safe, Persistent, and Evolving Agent Harness for Understanding Partially Observable Worlds&lt;/a&gt;&lt;small&gt;from the team of Yangqiu Song, HKUST; code released; a new benchmark · Actions of LLM agents in enterprise workflows must strictly comply with organizational policies, and the agents must handle hidden side effects and long-horizon state tracking under partial observability.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://doi.org/10.1109/csr69374.2026.11759868&quot;&gt;Constrained-Action AI Remediation for SIEM/XDR via a NeMo-Guardrails Proxy&lt;/a&gt;&lt;small&gt;code released; a new benchmark · When an LLM handles alerts automatically in a SOC, adversarial alerts can induce it, through LLM reasoning, to recommend and execute dangerous commands.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.10619v1&quot;&gt;TestJack: Should you trust the results in coding benchmarks? Agentic Coding Benchmarks Auditing via Evaluator Evolution&lt;/a&gt;&lt;small&gt;from the team of Dawn Song, University of California, Berkeley; a new benchmark · Coding benchmarks judge correctness only by fixed unit tests, so an agent can reward hack or miss requirements and still pass, which distorts scores.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2610.09462v1&quot;&gt;TMT: Runtime Backdoor Detection for Vision-Language-Action Policies on Unseen Tasks&lt;/a&gt;&lt;small&gt;code released; a new benchmark · Detecting, at runtime, backdoored VLA policies activated by a trigger, especially on unseen tasks.&lt;/small&gt;&lt;/li&gt;&lt;/ol&gt;&lt;p class=&quot;sub&quot;&gt;Chosen by rule: venue tier, a ranked team, released code, a new benchmark and use by later papers; at most two per theme.&lt;/p&gt;&lt;h2&gt;Direction radar&lt;/h2&gt;&lt;table&gt;&lt;tr&gt;&lt;th&gt;Direction&lt;/th&gt;&lt;th&gt;This period&lt;/th&gt;&lt;th&gt;Average before&lt;/th&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ VLA safety&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;1&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ Memory poisoning and long-term memory&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;1&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;6.7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ Tool-call authorisation and approval&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;8.7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ Indirect injection defences for tool agents&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;2&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ Reward hacking, self-improvement, research agents&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;2&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;6.7&lt;/td&gt;&lt;/tr&gt;&lt;/table&gt;&lt;h2&gt;Who is moving&lt;/h2&gt;&lt;p&gt;Institutions with their first paper in the field: International Hellenic University, AWS Forward Deployed Engineering, Corners Co., Cribl AI Research Lab, Department of Computational and Data Sciences, Department of Electrical and Systems Engineering, Digital Economy Lab, Northeastern University at Qinhuangdao, Pazhou Laboratory, Philadelphia&lt;/p&gt;&lt;h2&gt;Open source and models&lt;/h2&gt;&lt;ul&gt;&lt;li&gt;affaan-m/ECC &lt;small&gt;+4,644 stars, now 275,241&lt;/small&gt;&lt;/li&gt;&lt;li&gt;ifixai-ai/iFixAi &lt;small&gt;+4,333 stars, now 22,656&lt;/small&gt;&lt;/li&gt;&lt;li&gt;earendil-works/pi &lt;small&gt;+2,421 stars, now 113,498&lt;/small&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h2&gt;Which conclusions changed&lt;/h2&gt;&lt;table&gt;&lt;tr&gt;&lt;td&gt;Papers on the site&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;5,593 → &lt;b&gt;5,645&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;New papers in the last 12 months&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3,758 → &lt;b&gt;3,773&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;Institutions with a paper on the site&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3,568 → &lt;b&gt;3,585&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;Open-source repositories in the library&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;7,671 → &lt;b&gt;7,724&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;/table&gt;&lt;p class=&quot;sub&quot;&gt;The other 9 tracked numbers moved by less than half a point.&lt;/p&gt;&lt;h2&gt;Revision of the period before&lt;/h2&gt;&lt;p&gt;2026-09-27 – 2026-10-03 was first counted at 149 papers; 0 arrived later, 149 now.&lt;/p&gt;</content></entry><entry><id>https://xtra-agent-safety-atlas.pages.dev/issues/w-2026-10-03.en.html</id><title>Weekly issue 1 · 27 September – 3 October 2026</title><link href="https://xtra-agent-safety-atlas.pages.dev/issues/w-2026-10-03.en.html"/><updated>2026-10-05T02:30:00Z</updated>
<summary>Runtime guardrails and multi-turn trajectories had 8 new papers this week, against an average of 1.7 in the three before.</summary><content type="html">&lt;p class=&quot;head&quot;&gt;Runtime guardrails and multi-turn trajectories had 8 new papers this week, against an average of 1.7 in the three before.&lt;/p&gt;&lt;div class=&quot;tiles&quot;&gt;&lt;div&gt;&lt;b&gt;149&lt;/b&gt;&lt;span&gt;new papers&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;114.3&lt;/b&gt;&lt;span&gt;average first count of the three before&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;0&lt;/b&gt;&lt;span&gt;newly confirmed at a venue&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;134&lt;/b&gt;&lt;span&gt;new repositories&lt;/span&gt;&lt;/div&gt;&lt;div&gt;&lt;b&gt;13%&lt;/b&gt;&lt;span&gt;with an industry author&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;h2&gt;The main points&lt;/h2&gt;&lt;ol&gt;&lt;li&gt;&lt;b&gt;10 papers&lt;/b&gt; · Most new papers: VLA safety&lt;small&gt;Next: Memory poisoning and long-term memory 10; Tool-call authorisation and approval 10; Reward hacking, self-improvement, research agents 8&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;8 vs 1.7 before&lt;/b&gt; · Rising: Runtime guardrails and multi-turn trajectories&lt;small&gt;Also rising: Prompt injection detection and model-level defence 7; Agent skills ecosystem security 7; Safety evaluation of tool use, risk awareness 7&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;20 new citations&lt;/b&gt; · Most used earlier paper: Defeating Prompt Injections by Design&lt;small&gt;4 of them use it as a method, a benchmark or a comparison. 2025-03 · Indirect injection defences for tool agents&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;13%&lt;/b&gt; · of new papers have an author from industry&lt;small&gt;11% in the three periods before. 46 institutions published their first paper in the field.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;b&gt;+5,110 stars&lt;/b&gt; · Fastest-growing repository: NVIDIA/OpenShell&lt;small&gt;134 repositories were created in this period.&lt;/small&gt;&lt;/li&gt;&lt;/ol&gt;&lt;h2&gt;Worth reading first&lt;/h2&gt;&lt;ol&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2609.36308v1&quot;&gt;CheatBench: Measuring Reward Gaming in AI Agents&lt;/a&gt;&lt;small&gt;from the team of Dan Hendrycks, Center for AI Safety; code released; a new benchmark · RL-trained agents pursuing high reward may cheat (reward gaming), which does not match the user&amp;#x27;s intent and creates safety risks.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2609.33102v1&quot;&gt;ORBIT: A Framework for Multi-Agent Safety and Security Evaluations&lt;/a&gt;&lt;small&gt;from the team of Christian Schroeder de Witt, University of Oxford; code released; a new benchmark · Threats such as cascading prompt injection and inter-agent collusion in multi-agent LLM systems lack unified evaluation infrastructure.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2609.33628v2&quot;&gt;Climbing the Hill: Prompt Injection Red-Teaming Against Frontier Models with Curriculum Reinforcement Learning&lt;/a&gt;&lt;small&gt;from the team of Jinyuan Jia, Pennsylvania State University; code released · When training prompt injection attackers with RL to red-team frontier LLMs, attacks all fail and yield zero reward, so learning cannot get started (cold start).&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2609.36573v1&quot;&gt;CyberPersistBench: Evaluating LLM-Based Cyber Attackers on Installation and Persistence&lt;/a&gt;&lt;small&gt;from the team of Jing Shao, Shanghai Artificial Intelligence Laboratory; a new benchmark · Existing cybersecurity benchmarks stop at initial compromise and do not evaluate whether an LLM attacker can install and maintain a persistent foothold after intrusion.&lt;/small&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/abs/2609.38415v1&quot;&gt;Evaluating Whether GPT-6 Astra Performs Unsanctioned Supply-Chain Attacks&lt;/a&gt;&lt;small&gt;code released; a new benchmark · Whether frontier models, when working on difficult cybersecurity tasks, go beyond the task scope and launch supply chain attacks against third parties.&lt;/small&gt;&lt;/li&gt;&lt;/ol&gt;&lt;p class=&quot;sub&quot;&gt;Chosen by rule: venue tier, a ranked team, released code, a new benchmark and use by later papers; at most two per theme.&lt;/p&gt;&lt;h2&gt;Direction radar&lt;/h2&gt;&lt;table&gt;&lt;tr&gt;&lt;th&gt;Direction&lt;/th&gt;&lt;th&gt;This period&lt;/th&gt;&lt;th&gt;Average before&lt;/th&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Runtime guardrails and multi-turn trajectories&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;8&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;1.7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Prompt injection detection and model-level defence&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;7&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;1.7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Agent skills ecosystem security&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;7&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;1.7&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ Safety evaluation of tool use, risk awareness&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;7&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;2&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↑ VLA safety&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;10&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;5.3&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;↓ Coding agents and software engineering&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;1&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;4&lt;/td&gt;&lt;/tr&gt;&lt;/table&gt;&lt;h2&gt;Who is moving&lt;/h2&gt;&lt;p&gt;Institutions with their first paper in the field: Yinwang Intelligent Technology Co., APTA AI, Abundant AI, Adrenaline AI, BYD Company Limited, Centre for Health Informatics, Centre for Intelligent Cloud Computing, Cherkasy State Business College, Communication University of China, Corvic AI Research&lt;/p&gt;&lt;h2&gt;Open source and models&lt;/h2&gt;&lt;ul&gt;&lt;li&gt;NVIDIA/OpenShell &lt;small&gt;+5,110 stars, now 13,893&lt;/small&gt;&lt;/li&gt;&lt;li&gt;affaan-m/ECC &lt;small&gt;+3,880 stars, now 270,597&lt;/small&gt;&lt;/li&gt;&lt;li&gt;google/ax &lt;small&gt;+2,640 stars, now 12,777&lt;/small&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h2&gt;Which conclusions changed&lt;/h2&gt;&lt;table&gt;&lt;tr&gt;&lt;td&gt;Papers on the site&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;5,513 → &lt;b&gt;5,593&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;New papers in the last 12 months&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3,704 → &lt;b&gt;3,758&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;Institutions with a paper on the site&lt;/td&gt;&lt;td class=&quot;n&quot;&gt;3,540 → &lt;b&gt;3,568&lt;/b&gt;&lt;/td&gt;&lt;/tr&gt;&lt;/table&gt;&lt;p class=&quot;sub&quot;&gt;The other 10 tracked numbers moved by less than half a point.&lt;/p&gt;&lt;h2&gt;Revision of the period before&lt;/h2&gt;&lt;p&gt;2026-09-20 – 2026-09-26 was first counted at 121 papers; 1 arrived later, 122 now.&lt;/p&gt;</content></entry></feed>
