Astrolabe for Agent Safety · Periodic report
Monthly issue · October 2026
2026-10-01 – 2026-10-31 · published 2026-11-06 · 中文
Coding agents and software engineering had 34 new papers this month, against an average of 12.3 in the three before.
653new papers
466average first count of the three before
47newly confirmed at a venue
286new repositories
14%with an industry author
The main points
- 34 papers · Most new papers: Process evaluation and execution auditing of agentsNext: Coding agents and software engineering 34; Tool-call authorisation and approval 30; VLA safety 22
- 34 vs 12.3 before · Rising: Coding agents and software engineeringAlso 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
- 47 papers · Newly confirmed as accepted or publishedMost: ASE 2026 10; ICRA 2026 9; AAAI 2026 6
- 64 new citations · Most used earlier paper: $τ$-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains15 of them use it as a method, a benchmark or a comparison. 2024-06 · Process evaluation and execution auditing of agents
- 14% · of new papers have an author from industry11% in the three periods before. 292 institutions published their first paper in the field.
- +14,128 stars · Fastest-growing repository: affaan-m/ECC286 repositories were created in this period.
Worth reading first
- Compaction Laundering: Provenance Destruction in Context-Compacting Coding AgentsShows 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.
- Agentic Backdoors via Pretraining PoisoningFirst 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's machine.
- Red-Teaming Multi-Agent Attacks for Loss-of-Control RisksMiniLab 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.
- SecPO: Principled Adversarial Training for Prompt Injection SecurityAdversarial 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.
- My Credentials, Please: Detecting Filesystem Capability Laundering in MCP Servers at ScaleLarge-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).
- Approve What, Exactly? Effect Binding and Trusted Rendering for Agent Step-Up AuthorizationDefines 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.
- Training Natural Schemers: RL Can Amplify Oversight-Conditional BehaviorBuilds 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.
- Do Tool Calls Execute as Intended? Measuring and Repairing Intent-Execution Correspondence in LLM AgentsMeasures production sessions and finds harnesses silently alter tool calls: Claude Code's Bash tool changes 12% of code-bearing calls, often running the wrong action with no error reported. · When an LLM agent's tool calls are passed through multiple hops, they can be silently rewritten, causing wrong actions to be executed or pointless retries.
- Refuses in Chat, Complies with Arms: LLM Refusal Does Not Survive Robot ControlMatched 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.
- Caught in the Act: Probes Effectively Detect Sabotage and Catch Unverbalized DeceptionMulti-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.
The monthly star papers: two models (Claude Opus 5.5 and GPT-6.1 Sol) read every paper of the month and vote.
Direction radar
| Direction | This period | Average before |
|---|---|---|
| ↑ Coding agents and software engineering | 34 | 12.3 |
| ↑ Indirect prompt injection: toolkits, defence evaluations and accumulating attacks | 22 | 4.3 |
| ↑ Prompt injection detection and model-level defence | 22 | 8.3 |
| ↑ Contextual integrity and privacy leakage | 19 | 6.3 |
| ↑ Tool-call authorisation and approval | 30 | 19.3 |
| ↓ RAG poisoning and retrieval security | 5 | 12 |
| ↓ Agent identity, delegation, decentralised protocols | 4 | 10 |
Who is moving
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.
Open source and models
- affaan-m/ECC +14,128 stars, now 284,725
- ifixai-ai/iFixAi +12,268 stars, now 30,591
- earendil-works/pi +5,600 stars, now 116,677
Which conclusions changed
| Papers on the site | 5,593 → 6,318 |
| New papers in the last 12 months | 3,758 → 4,314 |
| Papers about digital (software) agents | 72.4% → 73.6% |
| Institutions with a paper on the site | 3,568 → 3,889 |
| Papers of the last 12 months that release code | 17.9% → 17.4% |
| Open-source repositories in the library | 7,671 → 7,889 |
| Papers using a language model that use at least one open-weight model | 75.4% → 76.4% |
The other 6 tracked numbers moved by less than half a point.
Revision of the period before
2026-09-01 – 2026-09-30 was first counted at 564 papers; 5 arrived later, 569 now.
Open this issue in the Astrolabe