Astrolabe for Agent Safety · Periodic report
Weekly issue 4 · 18 – 24 October 2026
2026-10-18 – 2026-10-24 · published 2026-11-06 · 中文
The theme “Alignment, oversight” had 16 new papers this week, against an average of 4.7 in the three before.
141new papers
101.7average first count of the three before
0newly confirmed at a venue
58new repositories
11%with an industry author
The main points
- 10 papers · Most new papers: Process evaluation and execution auditing of agentsNext: Tool-call authorisation and approval 8; Prompt injection detection and model-level defence 7; Coding agents and software engineering 6
- 5 vs 1.0 before · Rising: Human-agent autonomy, governance, trust frameworksAlso rising: CoT monitoring and steganography 5; AI control and sabotage evaluation 4; Agent societies and cooperation 5
- 5 new citations · Most used earlier paper: Representation Engineering: A Top-Down Approach to AI Transparency3 of them use it as a method, a benchmark or a comparison. 2023-10 · Deception, scheming and sandbagging
- 11% · of new papers have an author from industry14% in the three periods before. 66 institutions published their first paper in the field.
- +4,546 stars · Fastest-growing repository: affaan-m/ECC58 repositories were created in this period.
Worth reading first
- AutoRedClaw: Automated Red-Teaming for Claw-Style Agentsfrom 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.
- Towards Robust Model Organisms for Eliciting Latent Knowledgefrom 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.
- Act on Your Behalf, Beyond Your Authorization: Benchmarking LLM Agent Authorization Alignmentfrom 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.
- Humans Correct Their Judgments Through Debate, but Weaker AI Models Do NotNeurIPS 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.
- After the Attacker Leaves: Benchmarking Carryover Attacks on LLM AgentsEMNLP 2026 · Main conference (A-tier venue); a new benchmark · Whether a one-time indirect prompt injection, once written into an agent's persistent state, can be triggered repeatedly across sessions after the attacker has left.
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 |
|---|---|---|
| ↑ Human-agent autonomy, governance, trust frameworks | 5 | 1 |
| ↑ CoT monitoring and steganography | 5 | 1 |
| ↑ AI control and sabotage evaluation | 4 | 0.7 |
| ↑ Agent societies and cooperation | 5 | 2 |
| ↑ Coding agents and software engineering | 6 | 3 |
| ↓ Runtime guardrails and multi-turn trajectories | 0 | 4.3 |
| ↓ Reward hacking, self-improvement, research agents | 1 | 3.3 |
| ↓ Agent skills ecosystem security | 2 | 4 |
Who is moving
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
Open source and models
- affaan-m/ECC +4,546 stars, now 284,725
- ifixai-ai/iFixAi +2,313 stars, now 30,591
- earendil-works/pi +1,250 stars, now 116,677
Which conclusions changed
| Papers on the site | 5,781 → 6,318 |
| New papers in the last 12 months | 3,868 → 4,314 |
| Papers about digital (software) agents | 72.8% → 73.6% |
| Institutions with a paper on the site | 3,646 → 3,889 |
| Open-source repositories in the library | 7,792 → 7,889 |
| Papers using a language model that use at least one open-weight model | 75.3% → 76.4% |
The other 7 tracked numbers moved by less than half a point.
Revision of the period before
2026-10-11 – 2026-10-17 was first counted at 105 papers; 52 arrived later, 157 now.
Open this issue in the Astrolabe