Stop secret leaks from agents
A secret scan across prompts, tool calls and results, where it blocks, and what it cannot see.
The problem
Agents read configuration files, print tokens into logs and paste credentials into commands. Once a secret reaches a model prompt, a tool result or a pushed commit, it is already outside your control. A local scan blocks the movement before it completes.
What the action gate scans
The secret gate scans agent prompts, tool calls and tool results for credential shapes: API keys, tokens, private keys and cloud credentials. A tool result carrying a secret is withheld whole, with the reason naming the rule, never the secret text. Prompts and calls carrying secrets are refused before they run.
Pushes carrying secrets
The install-time secret scan blocks a push when it finds an exposed credential in the pushed commits. Fix the leak at its source, rotate the credential, and push again. The block names the finding class so you know what to search for.
Limits
Pattern matching cannot see every secret: credentials embedded in ordinary prose or unusual encodings can pass, and server-side logs redact fields by key name rather than by value. Treat the action gate as one layer: keep secrets out of repositories and prompts in the first place, rotate anything that may have leaked, and review account access after an incident.
References
- OWASP Top 10 for LLM Applications, OWASP GenAI Security Project
- AI Risk Management Framework, NIST
- Gitleaks secret scanner, Gitleaks