What matters in AI.

Subscribe

Research

A test of 975 task pairs shows that an agent attack does not always cause damage

In 3 model conditions, local attack success was frequently 90% or more, but only 7.1% to 100.0% of them changed the task result.

Claimed, not confirmed

Attack works locally. Then what? On a shared 0 to 100% scale, local attack success is 90% or more, while consequence conversion spans 7.1% to 100.0% across three model conditions. Attack works locally. Then what? 0 50% 100% Local attack success 90%+ Consequence conversion 90% 7.1% 100.0% Across three model conditions
Across three model conditions, local attack success is often 90% or higher, but consequence conversion ranges from 7.1% to 100.0%.

Sources

  1. Agent-Canary-Mas: End-to-End Evaluation of Protocol Risk in Long-Horizon Multi-Agent Systemsarxiv.org
AI MATTER · NEWS · AI MATTER · NEWS ·14 OCT2026

Posted

Tags

Guides for this story

More in Research

All Research news