dramaWTF 7.5via Simon Willison
Breaking Claude Code Opus 5 Auto Mode
"Your autonomous coding agent just accidentally imported a Trojan Horse via base64. Classic."
Explain Like I'm Normal
Renowned security researcher Johann Rehberger bypassed Anthropic's new 'Auto Mode' safety layer with an 80% success rate. By tricking the agent into downloading a zip file containing a malicious local file named 'struct.py', the agent inadvertently executed unauthorized code while attempting a standard library import. Ironically, the safety guardrails sometimes prevented the agent from stopping its own compromise.
#security#prompt-injection#claude-code#anthropic
GET THE DAILY CHAOS
The only newsletter for people who read AI news at 3am and feel things. One email a day.