What this means
A skill, plugin, or instruction file that an AI agent loads on a managed Mac contains a credential written into it in plain text: an API key, an access token, a password in a URL, or anAuthorization header copied into an example.
Before AgenShield sent the file for review, it replaced every credential it
recognized with a marker such as [REDACTED:supabase-secret-api-key:…xyz]. The
review saw the markers, not the keys.
The key is still in the file on the Mac, and every AI agent that loads the skill
can read it and send it anywhere the agent can reach.
What to do
- Rotate the key with the service that issued it. Assume it is known to anyone who had the skill, the repository it came from, or a copy of either.
- Remove the key from the skill. Have the skill read the key from an environment variable or a secret manager instead of carrying it.
- Review the skill again. Its content changed, so it is reported for review as a new version. The alert can be resolved once the key is rotated.