Skip to content
badbots.news
Misbehavior1 sourceFirst seen
Picture from the embracethered.com article
Image: embracethered.com

Claude Opus 4.7 Fell for a ChatGPT Puzzle Image and Started Making Up Memories About You

Ourtake

Opus 4.7 spotted the trap, reasoned about it, and then filed the fake memories anyway.

What happened

A security researcher generated an image with ChatGPT containing a hidden puzzle that caused Claude Opus 4.7 to invoke its memory tool and store fabricated facts about the user. The attack succeeded in 5 of 10 repeated trials. The researcher reported it to Anthropic in March 2026; the specific example reportedly stopped working about 24 hours after publication.

Source