AI agents can pass safety checks and still leak secrets through handoffs

A new assessment of Anthropic, Google, and OpenAI agent workflows found that secure-looking defaults can still expose secrets when one processing stage trusts another. The attacks did not require the agent to violate an obvious safety rule: approved commands, shared workspaces, and automatic output publishing combined into data-exfiltration paths.
Source

👉@sysadminoff

https://4sysops.com/archives/ai-agents-can-pass-safety-checks-and-still-leak-secrets-through-handoffs/