Astrolabe for Agent Safety · Periodic report

Weekly issue 6 · 1 – 7 November 2026

2026-11-01 – 2026-11-07 · published 2026-11-12 · 中文

The theme “Evaluation, guardrails” had 10 new papers this week, against an average of 5.7 in the three before.

127new papers
126.7average first count of the three before
191newly confirmed at a venue
26new repositories
16%with an industry author

The main points

  1. 6 papers · Most new papers: VLA safetyNext: Agentic system architecture and runtime engineering 6; Indirect injection defences for tool agents 5; Process evaluation and execution auditing of agents 5
  2. 5 vs 0.7 before · Rising: Prompt injection and information leakage in LLM-integrated applicationsAlso rising: Execution control and assurance in agent frameworks 5; Indirect injection defences for tool agents 5; MCP servers and protocol security 3
  3. 191 papers · Newly confirmed as accepted or published3 at the top-4 security or top-3 ML venues. Most: EMNLP 2026 134; Proceedings of the Second Workshop for REsearch on Agent Language Models (REALM 2026) 2026 7; AAAI 2026 6
  4. 16 new citations · Most used earlier paper: AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agents2 of them use it as a method, a benchmark or a comparison. 2024-10 · Safety evaluation of tool use, risk awareness
  5. 16% · of new papers have an author from industry13% in the three periods before. 62 institutions published their first paper in the field.
  6. +15,971 stars · Fastest-growing repository: affaan-m/ECC26 repositories were created in this period.

Worth reading first

  1. MetaPI: Constructing Prompt Injection Benchmarks from LLM Agent BenchmarkNeurIPS 2026 · Main conference (top-3 ML venue); from the team of Yuan Tian, University of California, Los Angeles; a new benchmark
  2. Failure Prediction and Classification in Vision-Language-Action Models Via Calibrated Latent Probingfrom the team of Ufuk Topcu, The University of Texas at Austin; code released; a new benchmark
  3. Airlock: Bounding the Multi-Agent Blast Radius via Selective Online Containmentfrom the team of Hamed Haddadi, Imperial College London; code released; a new benchmark · In multi-agent systems, a single agent's erroneous output spreads across tasks before it is detected, creating a system-level blast radius.
  4. MCP Tool Poisoning in Agentic AI Frameworks: A Cross-Framework Measurement Studycode released; a new benchmark · MCP feeds attacker-controllable tool descriptions, schema fields and response data into the model context without validation, leading to tool poisoning.
  5. The Customer Never Sees the Goal: Measuring and Detecting Vulnerability-Targeted Manipulation in LLM Assistantscode released; a new benchmark · An operator sets a business goal for an LLM assistant that customers cannot see, and the assistant may manipulate customers by exploiting vulnerabilities they disclose (VTM).

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
↑ Prompt injection and information leakage in LLM-integrated applications50.7
↑ Execution control and assurance in agent frameworks51.7
↑ Indirect injection defences for tool agents52.7
↑ MCP servers and protocol security31
↑ Agent red teaming and cross-session threats31
↓ Contextual integrity and privacy leakage04.3
↓ Prompt injection detection and model-level defence25.7
↓ Coding agents and software engineering36.7
↓ AI control and sabotage evaluation03.3

Who is moving

Institutions with their first paper in the field: Eastern International University, Jiffy Labs, Thu Dau Mot University, University of Information Technology, ABLEs, AGH University of Krakow, AWS Agentic AI, Abylkas Saginov Karaganda Technical University, Achiral Research, Acompany Co.

Open source and models

Which conclusions changed

Papers on the site6,318 → 6,421
New papers in the last 12 months4,314 → 4,397
Institutions with a paper on the site3,889 → 3,942
Open-source repositories in the library7,889 → 8,073

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

Revision of the period before

2026-10-25 – 2026-10-31 was first counted at 134 papers; 3 arrived later, 137 now.

Open this issue in the Astrolabe