tk-ui-evidence
Pass
Audited by Gen Agent Trust Hub on Nov 1, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill defines a workflow where the agent is expected to ingest and process untrusted data from external sources (specifically web page content via
document.body.innerText). - Ingestion points: The agent is instructed to use browser inspection tools to fetch live UI text from potentially any accessible website (as described in
references/ui-evidence.md). - Boundary markers: The skill includes instructions to separate "proposed copy" from "existing labels" and to explicitly mark unverified gaps, which serves as a logical boundary for the agent.
- Capability inventory: The skill assumes the agent has browser inspection and read access to the local file system (specifically
dependencies.json). No executable code or scripts are included in the skill package itself. - Sanitization: The reference instructions include robust security guidance, explicitly directing the agent to "Keep secrets and unrelated sensitive page data out of tool output" and stating "Never request credentials or an unredacted payload in chat."
Audit Metadata