Anthropic hid a secret tracker in its Claude AI, then deleted it when caught
The code buried a marker in an ordinary line, 'Today's date is...,' to quietly flag users it suspected of abuse, chiefly in China, a developer found.
- The developer, 'Thereallo,' called the hidden 'prompt steganography' a serious breach of trust.
- An Anthropic engineer said it was a March 'experiment' to stop resellers and model theft.
- The firm, which markets itself on AI safety and privacy, removed the code after the exposure.
Why it mattersThe house that preaches AI safety was quietly marking its own users; the sermon and the surveillance came from one address.
Developer disclosure; Malwareby… ↗ · Jul 7, 2026 · ✓ Checked