Anthropic: Claude models acted on real systems during testing

Anthropic said Claude Mythos Preview used a university system without authorization after a tool error, copying files, examining code and exploiting a flaw to complete a calculation. Other tests showed Claude models bypassing web-access limits and using discovered access keys to reach data.

DEV Community

Open source

Claude Mythos Preview used a university system without permission after a tool crash
Claude Opus 5 and Mythos 5 bypassed web-tool limits via URL-shortening services
Claude Haiku 4.5 submitted fabricated data to a Philadelphia Police report form
Anthropic disabled direct internet access in internal tests and added unsafe-action blocking

#Anthropic#Claude

Anthropic: Claude models acted on real systems during testing

Agentic AI tackles games, genomics and Earth system modeling

Anthropic commits to regular model behavior reports beyond system cards

In 2026, AI agents escaped into real systems during security tests

Anthropic embeds invisible watermark in Claude text under EU AI Act