agent-canary

This Skill addresses the challenge of proactively detecting malicious skills that steal credentials within an AI Agent workspace. In real-world development, agent skills may be embedded with malicious code that silently reads or copies sensitive information such as environment variables and key files. Traditional security tools often rely on passive log auditing and cannot identify such activities in real time.

Usage is straightforward: simply tell the AI "deploy decoys" or "set up honeypot," and it will automatically plant multiple files disguised as real credentials (e.g., .env.canary, secrets.backup.json) in the workspace root, memory directory, and user directories. These files contain tokens marked with a "CANARY" identifier. If any skill reads, copies, or exfiltrates these files, the monitoring script triggers an immediate alert. You can then check the status with "canary check" and clean up all decoys with "canary cleanup."

It is ideal for teams or individuals who need to secure their AI development environments, especially those using platforms like OpenClaw to manage multiple agent skills and are concerned about backdoors in third-party skills. Operations engineers can use it to quickly verify if skills are accessing sensitive data beyond their permissions, and agent developers can employ it as a supplementary tool for security testing.

We recommend using this tool when deploying new skills for the first time or during regular security checks. Note that it primarily targets file system access behaviors and cannot detect data exfiltration at the network level. Additionally, the decoy files contain obvious identifiers and should only be used in test environments; they are not a substitute for formal security auditing processes.