Ship Safe 10.0.0 Adds Security Review Layer for Hermes AI Agent Tool Access

Hermes Agent, an AI system capable of running tools, connecting to MCP servers, and scheduling tasks, introduces a broader security challenge by extending the attack surface beyond simple prompt safety checks. The core concern is tracing the full path an untrusted input can travel—from a pull request or README through agent context, permissions, and credentials to an external action. Ship Safe 10.0.0 addresses this by adding an investigation layer that derives evidence-backed verdicts—confirmed, likely, unresolved, or refuted—rather than flagging unproven patterns as definitive vulnerabilities. The new version covers Hermes-specific areas including cron lifecycle, credential paths, MCP initialization, and terminal backend posture, with findings exportable to JSON and SARIF for CI integration. A key vulnerability demonstrated in a test fixture shows that update_job applies only a non-empty check when merging new payloads, unlike the stricter lifecycle check in create_job, allowing a changed payload to persist and later be dispatched without equivalent scrutiny.