light-idea-critique
Pass
Audited by Gen Agent Trust Hub on Jul 17, 2026
Risk Level: SAFE
Full Analysis
- [PROMPT_INJECTION]: The skill implements a strict 'IRON RULE' (detailed in
SKILL.mdandreferences/protocol.md) that treats all user-provided ideas and external search results as data rather than instructions. It specifically monitors for and flags common injection patterns such as 'ignore previous instructions' or requests to manipulate scoring, maintaining a read-only posture regarding the analyzed content. - [DATA_EXFILTRATION]: The skill performs network operations to fetch academic metadata from well-known services, including OpenAlex (
api.openalex.org) and Semantic Scholar (api.semanticscholar.org). These operations are strictly aligned with the skill's primary purpose of novelty verification. No sensitive local file access or unauthorized data transmission was observed. - [COMMAND_EXECUTION]: The skill utilizes several local Python scripts (e.g.,
score_aggregate.py,novelty_audit.py,sycophancy_guard.py) to perform scoring, novelty auditing, and metrics calculation. These scripts are self-contained, use standard libraries, and do not execute untrusted code or system commands beyond their documented utility functions. - [INDIRECT_PROMPT_INJECTION]: As an idea critique tool, the skill naturally ingests untrusted user data. It effectively mitigates the risk of indirect prompt injection by using physical information isolation protocols (Phase 1 BLIND vs Phase 2 OPEN) and explicit boundary markers. The capability inventory is limited to standard API calls and local data processing, with robust sanitization logic established in the protocol.
Audit Metadata