Developer finds silent approval gate bypass in AI agent tool, builds attack suite to fix it

A developer building an AI-powered incident response agent called sentinel-agent discovered a critical flaw in TrueForge, an open-source agent harness, where tools lacking annotation metadata could bypass approval gates entirely and execute production actions without human sign-off. The bug stemmed from a four-line function that returned false for all permission checks when tool annotations were undefined, meaning an unannotated rollback tool would silently fire against production systems. The developer also found that the MCP server was bound to an unauthenticated endpoint, allowing requests to reach production tools while completely bypassing the harness-level approval gate. To address both issues, the developer implemented multiple fixes and then built a dedicated test suite designed solely to attack and stress-test those same safeguards. The project highlighted a broader engineering challenge: enforcing the boundary between automated investigation and human-authorised execution in AI agent systems.