Astrolabe for Agent Safety · Periodic report

Weekly issue 5 · 25 – 31 October 2026

2026-10-25 – 2026-10-31 · published 2026-11-06 · 中文

The theme “Tool execution, coding and runtime” had 40 new papers this week, against an average of 26.7 in the three before.

134new papers
99average first count of the three before
0newly confirmed at a venue
30new repositories
14%with an industry author

The main points

  1. 9 papers · Most new papers: Coding agents and software engineeringNext: VLA safety 6; Prompt injection detection and model-level defence 5; Process evaluation and execution auditing of agents 5
  2. 9 vs 4.7 before · Rising: Coding agents and software engineeringAlso rising: Reward hacking, self-improvement, research agents 5; VLA safety 6; Safety evaluation of tool use, risk awareness 5
  3. 11 new citations · Most used earlier paper: Defending Against Indirect Prompt Injection Attacks With Spotlighting5 of them use it as a method, a benchmark or a comparison. 2024-03 · Prompt injection detection and model-level defence
  4. 14% · of new papers have an author from industry13% in the three periods before. 65 institutions published their first paper in the field.
  5. 30 repositories · created in this period

Worth reading first

  1. Appreciation Teaches Unauthorized Execution: Self-Evolution Degrades the Robustness of Personal Agents to Indirect Prompt Injectionfrom the team of Chaowei Xiao, Johns Hopkins University; code released · Whether a self-evolving personal agent becomes more vulnerable to indirect prompt injection attacks after learning from user feedback.
  2. How Much Can Tool Calls Reveal About You? Privacy Leakage from Tool-Call Tracesfrom the team of Dongrui Liu, Shanghai Artificial Intelligence Laboratory; code released; a new benchmark · Whether MCP tool-call traces leak user privacy, allowing an observer to infer user profile attributes.
  3. SecPO: Principled Adversarial Training for Prompt Injection Securityfrom the team of David Wagner, University of California, Berkeley; code released · LLM agents have insufficient defenses against prompt injection; the strongest defense, Meta-SecAlign, has an ASR close to 100% under adaptive attacks.
  4. Agentic Backdoors via Pretraining Poisoningfrom the team of Nicholas Carlini, Anthropic (United States); code released · 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.
  5. Your Agent Acts Differently When You're Around: How User Presence Changes Agent Behavior in Negotiationfrom the team of He He, NYU; a new benchmark · How does the behavior of an LLM agent negotiating on a user's behalf change when the user is in the loop, and does misaligned behavior appear?

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
↑ Coding agents and software engineering94.7
↑ Reward hacking, self-improvement, research agents51
↑ VLA safety62.3
↑ Safety evaluation of tool use, risk awareness51.7
↑ Human oversight and control of agents30.3

Who is moving

Institutions with their first paper in the field: Cygnux Labs, Manchester Metropolitan University, ACE Robotics, ADAPT Research Unit CTTC-CERCA, AI Center, AIR Institute, Al-Ahliyya Amman University, American University of Sharjah, Anhui Polytechnic University, Baba

Open this issue in the Astrolabe