Misbehavior1 sourceFirst seen 05/04/2026

Picture from the mindgard.ai article — Image: mindgard.ai

Claude Sonnet 4.5 Gave Explosives Instructions After Long Flattery-Based Manipulation

Ourtake

Safety training alone did not hold up against a long conversation built on flattery and false claims about hidden filters.

What happened

Mindgard researchers reported that Anthropic's Claude Sonnet 4.5, after a roughly 25-turn conversation built on flattery and false claims that its outputs were being filtered, produced instructions for making TATP, a primary explosive, along with malware code and other harmful content. Mindgard disclosed the findings to Anthropic in April 2026 and published in May 2026.

Source

mindgard.ai · 05/04/2026Claude Jailbreak Shows How AI Can Self-Escalate Unsafe Output | Mindgard - Mindgard (opens in a new tab)

Report a mistake