The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides numerous pre-written bash and python commands intended to be executed in the agent's environment, such as aws ssm get-parameter, op read, and python -m relay.testbed. While these are for a specific testing workflow, they involve executing shell commands with potentially sensitive side effects.
[CREDENTIALS_UNSAFE]: The skill provides direct instructions and commands for retrieving production credentials (API keys, usernames, passwords) from secure vaults like AWS Systems Manager (SSM) and 1Password (op read). It explicitly references 1Password items such as op://DeLoSecrets/Gorilla Desk/username and `op://DeLoSecrets/IPM Surface approver
automaticai`.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external CRM data and JSON board files (~/code/james-brennan/boards/*.json).
Ingestion points: Reads job and customer data from the GorillaDesk API and local JSON board files.
Boundary markers: None mentioned for data processing, though a "geographic fence" is used to limit the scope of writes.
Capability inventory: Execution of shell commands, network requests to APIs via relay.testbed, and file system access to ~/code/james-brennan/.
Sanitization: None described for the content of the JSON fields being processed.
[PRIVILEGE_ESCALATION]: The instructions include unset AWS_ACCESS_KEY_ID AWS_SECRET_ACCESS_KEY AWS_SESSION_TOKEN to ensure that SSO-derived credentials stored in the environment are used instead of stale environment variables, effectively managing how the agent assumes its granted permissions.

ipm-live-testbed