Astrolabe for Agent Safety · Periodic report

Weekly issue 2 · 4 – 10 October 2026

2026-10-04 – 2026-10-10 · published 2026-10-12 · 中文

51 new papers this week, 62% fewer than the three before at their first count (133 on average).

51new papers
132.7average first count of the three before
11newly confirmed at a venue
91new repositories
18%with an industry author

The main points

  1. 5 papers · Most new papers: Process evaluation and execution auditing of agentsNext: Deception, scheming and sandbagging 4; Injection and deception against GUI and web agents 3; Prompt injection detection and model-level defence 3
  2. 11 papers · Newly confirmed as accepted or publishedMost: Proceedings of the ACM on Software Engineering 2026 3; SOSP 2026 1; Logical Methods in Computer Science 2026 1
  3. 2 new citations · Most used earlier paper: EIA: ENVIRONMENTAL INJECTION ATTACK ON GENERALIST WEB AGENTS FOR PRIVACY LEAKAGE2 of them use it as a method, a benchmark or a comparison. 2024-09 · Injection and deception against GUI and web agents
  4. 18% · of new papers have an author from industry13% in the three periods before. 16 institutions published their first paper in the field.
  5. +4,644 stars · Fastest-growing repository: affaan-m/ECC91 repositories were created in this period.

Worth reading first

  1. Formal Runtime Verification for Tool-Using LLM Agents: An Offline Same-Benchmark Study on AgentDojo and STACACM 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.
  2. Safe, Persistent, and Evolving Agent Harness for Understanding Partially Observable Worldsfrom 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.
  3. Constrained-Action AI Remediation for SIEM/XDR via a NeMo-Guardrails Proxycode 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.
  4. TestJack: Should you trust the results in coding benchmarks? Agentic Coding Benchmarks Auditing via Evaluator Evolutionfrom 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.
  5. TMT: Runtime Backdoor Detection for Vision-Language-Action Policies on Unseen Taskscode released; a new benchmark · Detecting, at runtime, backdoored VLA policies activated by a trigger, especially on unseen tasks.

Chosen by rule: venue tier, a ranked team, released code, a new benchmark and use by later papers; at most two per theme.

Direction radar

DirectionThis periodAverage before
↓ VLA safety17
↓ Memory poisoning and long-term memory16.7
↓ Tool-call authorisation and approval38.7
↓ Indirect injection defences for tool agents27
↓ Reward hacking, self-improvement, research agents26.7

Who is moving

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

Open source and models

Which conclusions changed

Papers on the site5,593 → 5,645
New papers in the last 12 months3,758 → 3,773
Institutions with a paper on the site3,568 → 3,585
Open-source repositories in the library7,671 → 7,724

The other 9 tracked numbers moved by less than half a point.

Revision of the period before

2026-09-27 – 2026-10-03 was first counted at 149 papers; 0 arrived later, 149 now.

Open this issue in the Astrolabe