infra-deploy
Pass
Audited by Gen Agent Trust Hub on Oct 20, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill defines a 'Self-Evolving' mechanism where the agent is instructed to modify the
SKILL.mdfile to fix errors or update procedures based on the results of command execution. This creates a vulnerability where malicious or malformed command output could influence the agent to write unsafe or persistent instructions back into the skill definition. - Ingestion points: Command output results from tools like
gcloud,docker, andcurlare used during the 'Reflection' phase to determine how to edit the skill. - Boundary markers: None identified; there are no delimiters protecting the self-editing process from adversarial input contained in execution feedback.
- Capability inventory: The skill uses
Bashfor command execution andWriteto modify its own source file. - Sanitization: No sanitization or validation logic is provided for the agent's self-modification process.
- [COMMAND_EXECUTION]: The skill relies heavily on the
Bashtool to perform administrative operations, including building Docker containers, pushing to registries, and deploying services to GCP Cloud Run. - [DATA_EXFILTRATION]: Secrets retrieved via the 1Password CLI (
op) are passed as environment variables ingcloud run deployand as query string parameters incurlrequests (e.g., to the Cal.com API). This methodology can expose sensitive data in process lists, system logs, or network metadata. - [EXTERNAL_DOWNLOADS]: The skill invokes
npx prisma migrate deployto fetch and execute the Prisma CLI from a remote registry. While this is standard for the workflow, it represents a runtime external dependency execution point.
Audit Metadata