← News

Agents & MCPSources: arXiv · 18 Nov 2026

Evolution Under Fire evolves agent safety harnesses against adaptive attacks

An arXiv paper proposes evolving executable safety harnesses for tool-using language-model agents, pairing harness evolution with adaptive stressors that preserve attacker utility while probing defenses.

In plain words

AI agents that use tools must handle content from untrusted sources like emails and webpages, where malicious instructions can hide alongside legitimate task directions. The paper's framework automatically evolves executable safety harnesses, the code layer that filters what an agent does, instead of relying on hand-designed static defenses. It also evolves attacks that keep tasks useful, so the defenses are tested against realistic pressure.

What was announced

The authors describe a staged framework in which utility-constrained, Pareto-scheduled tree search evolves complete executable safety harnesses from execution evidence, while the task model, benchmark, tools, and evaluator stay fixed. A second stage evolves adaptive stressors that discover utility-preserving attacks, and those attacks then guide further harness evolution. The approach targets the problem that existing defenses are largely hand-designed and static, so improvements depend on repeated manual redesign.

Reported results are limited to two benchmark-model pairings. On AgentDyn with GPT-4o-mini, an evolved harness reportedly cut attack success from 50.4% to 3.0% and improved clean and attacked utility over the initial harness. On Agent-SafetyBench with Qwen2.5-7B-Instruct, attack success fell from 82.6% to 62.0% while helpfulness and runtime validity were retained. Harnesses also transferred without further evolution to Qwen3.5-35B-A3B, reducing attack success from 32.1% to 0.9%. These figures come from the paper's abstract alone and have not been independently confirmed.

What it does not show

  • Evidence is limited to the paper's abstract; full methodology, experimental detail, and peer review status are not available from this source.
  • Reported gains cover only AgentDyn with GPT-4o-mini and Agent-SafetyBench with Qwen2.5-7B-Instruct, plus a transfer test on Qwen3.5-35B-A3B; no other models or settings are covered.

For security and IT teams

Scope
The source is the paper's abstract only, covering prompt-injection-style threats to tool-using agents that receive untrusted emails, webpages, and documents; it does not describe deployment guidance or availability.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.