TLDR Tech / InfoSec
Inside Claude Managed Agents: Reverse Engineering the Security Boundaries of Anthropic's Hosted Agent Runtime
Pluto Security reverse-engineered Anthropic's Claude Managed Agents hosted-agent runtime to probe its security features and limitations. Pluto found that the runtime uses the same gVisor sandbox as Cloud Cowork, the vault credential proxy provides strong secrets isolation, and network proxy bypasses utilize a multi-layer defense to prevent them. Pluto also noted that while there are strong security configurations, the default configuration is very permissive, and teams should ensure agents are secured thoroughly.
Read the full story