study-design-verify
Pass
Audited by Gen Agent Trust Hub on Aug 30, 2026
Risk Level: SAFE
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs the agent to perform read-only queries against live systems (databases, APIs, schemas) to verify how a system actually works. This is the primary purpose of the skill and is constrained by strict 'read-only' rules.
- [DATA_EXFILTRATION]: The skill requires reading system metadata, code, and usage statistics to provide advice. While this involves sensitive business data, the skill explicitly directs that raw query output be saved to a local evidence folder (
tmp/advice-study/) for human audit and does not suggest sending this data to external untrusted domains. Findings are summarized into structured digests. - [REMOTE_CODE_EXECUTION]: There are no patterns of downloading or executing external scripts. The skill relies on standard tools (likely via a Workflow tool in the agent platform) to perform queries and read files.
- [INDIRECT_PROMPT_INJECTION]: The skill recognizes the risk of processing untrusted data ('plausible-but-wrong is THE failure mode'). It mitigates this via an 'Adversarial Feasibility' phase where a separate agent re-verifies primary sources to catch errors or injections present in the study digest.
- [SAFE]: The skill includes extensive 'Red Flags' and 'Principles' sections that explicitly forbid writing to live systems and mandate verification of documentation against reality, which are excellent security and reliability best practices.
Audit Metadata