# Preprint reports stronger automated attacks across multi-turn AI conversations

_Published Monday, October 26, 2026 at 3:35 PM EDT · AI, Security, Science · Latest · Tier 2 — Notable_

A preprint introduces M³, an automated system for testing whether conversations can progressively bypass AI safety mechanisms. Researchers report that it outperformed leading multi-turn testing methods across two benchmarks and six target models, with attack success rate improvements of up to 53% on the ASR@5 measure.

The system pairs an attacker language model with another model that maintains a compact representation of the conversation's state. Even with a lightweight attacker, it reported gains of up to 26% on ASR@1 and 34% on ASR@5 against baselines using higher-capacity attackers. Those comparisons remain benchmark results reported in a preprint.

## Sources

- [cs.AI updates on arXiv.org](https://arxiv.org/abs/2610.31937)

---
Canonical: https://techandbusiness.org/newswire/0ocTg9tbnWPXRKx6vwDqFi
Published: 2026-10-26T19:35:56.680Z
Story chronology: 2026-10-26T04:00:00.000Z
Retrieved: 2026-10-26T23:00:21.233Z
Publisher: Tech & Business (techandbusiness.org)