eval-novel-style
Pass
Audited by Gen Agent Trust Hub on Sep 2, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted text (novel drafts) and interpolates them into analysis prompts. While the intent is stylistic evaluation, malicious instructions embedded in a draft could potentially influence the agent's output or cause it to ignore safety guidelines during the requested 'rewrite' phase.
- Ingestion points: The
SKILL.mdfile specifies reading 'draft text/path' for evaluation. - Boundary markers: No specific delimiters or 'ignore instructions' warnings are mentioned in
SKILL.mdorprompts/style_evaluator.mdwhen handling the draft text. - Capability inventory: The skill uses
scripts/novel_style_runtime_loader.pyandscripts/novel_style_eval.pyfor deterministic analysis and generates agent-driven reports and rewrites. - Sanitization: There is no evidence of input sanitization or filtering of the user-provided draft before it is processed by the LLM for evaluation or rewriting.
Audit Metadata