evidence-led-interviewer

Pass

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONTIME_DELAYED_CONDITIONALEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process user-provided career evidence, experience summaries, and interview responses, creating an attack surface for indirect instructions hidden in candidate data.
  • Ingestion points: The interaction loop described in SKILL.md for 'Knowledge discovery' and 'Formal evaluation' modes.
  • Boundary markers: The skill provides extensive rules to the agent, such as 'Prohibited inferences' in SKILL.md and 'Stop rules' in references/question-method.md, which serve to limit the agent's susceptibility to manipulative input.
  • Capability inventory: The skill records results to a database schema defined in references/record-contract.md and can route findings to the output-career-portfolio skill.
  • Sanitization: The 'Question workflow' and 'Behavioural anchors' require the agent to map inputs to specific job criteria and pre-defined anchors, reducing the likelihood of arbitrary user input overriding the agent's logical flow.
  • [TIME_DELAYED_CONDITIONAL]: The skill contains logic to perform a network-based version check only once every four hours, gating a potentially observable network operation behind a time-based cache.
  • Evidence: 'use its four-hour cached, read-only version check once at the start of a task' in SKILL.md.
  • [EXTERNAL_DOWNLOADS]: The skill attempts to connect to a network to check for updates to a companion skill if permitted by the user's environment.
  • Evidence: The reference to a version check for the 'why-hire-me-update' skill at the start of tasks in SKILL.md.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 09:17 PM
Security Audit — agent-trust-hub — evidence-led-interviewer