improve-prompt-deep
Pass
Audited by Gen Agent Trust Hub on Aug 14, 2026
Risk Level: SAFENO_CODE
Full Analysis
- [PROMPT_INJECTION]: The skill implements a defensive strategy against indirect prompt injection by treating all external data as untrusted. Evidence Chain: 1. Ingestion points: Web research results (references/research-policy.md, Section 4). 2. Boundary markers: Explicit instructions to ignore embedded requests (Section 4). 3. Capability inventory: File-write (Section 8). 4. Sanitization: Instruction filtering and path validation.
- [DATA_EXFILTRATION]: It enforces a mandatory data minimization step (Section 2, references/research-policy.md) that generalizes PII, secrets, and internal hostnames before external queries are made, effectively preventing data leakage.
- [SAFE]: File persistence is protected by robust path validation (Section 8, references/research-policy.md) that rejects path traversal, symlinks, and absolute paths, ensuring research notes are written only to authorized directories.
- [NO_CODE]: The skill contains only instructional documentation and configuration files; it does not provide or rely on executable scripts or binary dependencies.
Audit Metadata