MindSearch hit by critical code-injection flaw

This is a severe trust-boundary failure in an agent framework: model-generated code is treated as executable server code without authentication or sandboxing.

• –Public exploit details demonstrate command execution inside default deployments
• –Exposed API instances may leak environment secrets, alter files, or enable lateral movement
• –Docker configurations running as root substantially amplify impact
• –Operators should isolate port 8002, add authentication, and disable unsandboxed `exec()`
• –The incident underscores why agent frameworks need structured tool calls instead of free-form code execution