What matters in AI.

Subscribe

AutoPI finds more methods to break AI agents than other test tools

Against Claude Haiku 5.5, a strong target, AutoPI had 20% to 50% more attack success than recent tools.

Researchers show AutoPI, a tool that attacks AI agents with planted commands to find weak points. It aims to find many different methods to break a target, and not only to break it many times. On a test set, it broke the open model Qwen3.6-35B in all 6 classes of attack. The team releases its code and its attacks.

How it works

Most red teamers are tools that attack a system to test it. They use a small number of methods that work. AutoPI uses PIAtlas, a set of attack methods. An attacker model writes one attack text for each sub method. A judge model gives feedback, and the attacker changes the attacks that do not work.

The numbers

It broke one open model in all 6 classes of attack. Against Claude Haiku 5.5, a strong target, it had 20% to 50% more success and 25% to 400% more class coverage. It does not use training.

What is not known

The source gives no data on cost for each test. It does not say if the vendors of the targets will use the attacks to make their models strong.

Insight: One attack success rate hides which weak points remain

A public challenge found a 32% attack success rate on Meta SecAlign, a model with training against prompt injection. A rate counts how many times a target breaks, and not how many methods. AutoPI changes the score to coverage, which seems to turn a test from one number into a map of untested weak points. A vendor that gives only one rate does not show them.

Sources

  1. AutoPI: Taxonomy-Guided Adaptive Red Teaming for Prompt Injection in LLM Agentsarxiv.org
AI MATTER · NEWS · AI MATTER · NEWS ·19 NOV2026

Posted

Tags

Learn the terms in this story

Guides for this story

More in Security

All Security news