Astrolabe for Agent Safety · Periodic report
Weekly issue 1 · 27 September – 3 October 2026
2026-09-27 – 2026-10-03 · published 2026-10-05 · 中文
Runtime guardrails and multi-turn trajectories had 8 new papers this week, against an average of 1.7 in the three before.
The main points
- 10 papers · Most new papers: VLA safetyNext: Memory poisoning and long-term memory 10; Tool-call authorisation and approval 10; Reward hacking, self-improvement, research agents 8
- 8 vs 1.7 before · Rising: Runtime guardrails and multi-turn trajectoriesAlso rising: Prompt injection detection and model-level defence 7; Agent skills ecosystem security 7; Safety evaluation of tool use, risk awareness 7
- 20 new citations · Most used earlier paper: Defeating Prompt Injections by Design4 of them use it as a method, a benchmark or a comparison. 2025-03 · Indirect injection defences for tool agents
- 13% · of new papers have an author from industry11% in the three periods before. 46 institutions published their first paper in the field.
- +5,110 stars · Fastest-growing repository: NVIDIA/OpenShell134 repositories were created in this period.
Worth reading first
- CheatBench: Measuring Reward Gaming in AI Agentsfrom 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's intent and creates safety risks.
- ORBIT: A Framework for Multi-Agent Safety and Security Evaluationsfrom 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.
- Climbing the Hill: Prompt Injection Red-Teaming Against Frontier Models with Curriculum Reinforcement Learningfrom 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).
- CyberPersistBench: Evaluating LLM-Based Cyber Attackers on Installation and Persistencefrom 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.
- Evaluating Whether GPT-6 Astra Performs Unsanctioned Supply-Chain Attackscode 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.
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 |
|---|---|---|
| ↑ Runtime guardrails and multi-turn trajectories | 8 | 1.7 |
| ↑ Prompt injection detection and model-level defence | 7 | 1.7 |
| ↑ Agent skills ecosystem security | 7 | 1.7 |
| ↑ Safety evaluation of tool use, risk awareness | 7 | 2 |
| ↑ VLA safety | 10 | 5.3 |
| ↓ Coding agents and software engineering | 1 | 4 |
Who is moving
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
Open source and models
- NVIDIA/OpenShell +5,110 stars, now 13,893
- affaan-m/ECC +3,880 stars, now 270,597
- google/ax +2,640 stars, now 12,777
Which conclusions changed
| Papers on the site | 5,513 → 5,593 |
| New papers in the last 12 months | 3,704 → 3,758 |
| Institutions with a paper on the site | 3,540 → 3,568 |
The other 10 tracked numbers moved by less than half a point.
Revision of the period before
2026-09-20 – 2026-09-26 was first counted at 121 papers; 1 arrived later, 122 now.
Open this issue in the Astrolabe