douban-movie-review
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection attack surface.
- Ingestion points: External text data is ingested from public web pages (such as movie reviews on douban.com) when
agent.browser.execute_taskis executed inscripts/browser-use.py. - Boundary markers: Absent. There are no explicit delimiters or strict instructions telling the agent to ignore malicious instructions embedded within the retrieved web content.
- Capability inventory: The skill interacts with a remote browser session (
image_id="browser_latest") managed by theagentbaySDK. It does not possess any dangerous local file-write or subprocess execution privileges that could be directly targeted by a web payload. - Sanitization: Content pulled from the external browser session is evaluated directly without sanitization or structural validation.
Audit Metadata