Astrolabe for Agent Safety · Periodic report
Weekly issue 3 · 11 – 17 October 2026
2026-10-11 – 2026-10-17 · published 2026-10-19 · 中文
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).
The main points
- 8 papers · Most new papers: Process evaluation and execution auditing of agentsNext: Agentic system architecture and runtime engineering 7; Tool-call authorisation and approval 6; Prompt injection detection and model-level defence 5
- 7 vs 1.7 before · Rising: Agentic system architecture and runtime engineeringAlso rising: Coding agents and software engineering 5; Safety and alignment in reinforcement learning for language agents 3; Contextual integrity and privacy leakage 4
- 36 papers · Newly confirmed as accepted or publishedMost: ASE 2026 10; ICRA 2026 9; AAAI 2026 6
- 10 new citations · Most used earlier paper: $τ$-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains5 of them use it as a method, a benchmark or a comparison. 2024-06 · Process evaluation and execution auditing of agents
- 13% · of new papers have an author from industry15% in the three periods before. 38 institutions published their first paper in the field.
- +5,622 stars · Fastest-growing repository: ifixai-ai/iFixAi66 repositories were created in this period.
Worth reading first
- TRAP: Tool-result Replay Attack for Probing Misbinding in LLM Agentsfrom 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).
- LatentRiskBench: Measuring State-Dependent Risks in Enterprise AgentsEMNLP 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.
- AgentCIScope: Grounded Contextual Integrity Evaluation for LLM Agentsfrom 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.
- Scalable Weak-to-Strong Verifier Hardening for Agent Environmentscode 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.
- CoRedSkill: Stealthy Skill Composition Attacks via AST-Guided Injectionfrom 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.
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
| Direction | This period | Average before |
|---|---|---|
| ↑ Agentic system architecture and runtime engineering | 7 | 1.7 |
| ↑ Coding agents and software engineering | 5 | 2.7 |
| ↑ Safety and alignment in reinforcement learning for language agents | 3 | 0.7 |
| ↑ Contextual integrity and privacy leakage | 4 | 1.7 |
| ↑ Runtime authority and injection diagnosis | 3 | 1 |
| ↓ Reward hacking, self-improvement, research agents | 0 | 5.7 |
| ↓ Embodied task-planning safety | 0 | 4 |
| ↓ Agent societies and cooperation | 0 | 3.7 |
Who is moving
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
Open source and models
- ifixai-ai/iFixAi +5,622 stars, now 28,278
- affaan-m/ECC +4,938 stars, now 280,179
- earendil-works/pi +1,929 stars, now 115,427
Which conclusions changed
| Papers on the site | 5,645 → 5,781 |
| New papers in the last 12 months | 3,773 → 3,868 |
| Institutions with a paper on the site | 3,585 → 3,646 |
| Open-source repositories in the library | 7,724 → 7,792 |
The other 9 tracked numbers moved by less than half a point.
Revision of the period before
2026-10-04 – 2026-10-10 was first counted at 51 papers; 26 arrived later, 77 now.
Open this issue in the Astrolabe