sd-advanced-search
Pass
Audited by Gen Agent Trust Hub on Sep 30, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONPRIVILEGE_ESCALATION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from ScienceDirect (titles, abstracts, and metadata). While the data source is legitimate, the lack of sanitization and the presence of browser interaction capabilities creates a vulnerability to instructions embedded in external content.
- Ingestion points: Step 4 of SKILL.md.
- Boundary markers: None.
- Capability inventory: navigate_page, evaluate_script, and click tools.
- Sanitization: None.
- [PRIVILEGE_ESCALATION]: The skill explicitly instructs the agent to circumvent anti-automation measures on the target website. This includes modifying browser properties to hide automated status (webdriver) and automating the interaction with Cloudflare Turnstile captchas through cross-origin iframe manipulation.
Audit Metadata