agile-v-human-oversight

Pass

Audited by Gen Agent Trust Hub on Sep 13, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill establishes a framework for ingesting and processing data from multiple sources, including human precommits and outputs from other agents, which creates a potential surface for indirect prompt injection. However, it contains robust defensive instructions specifically targeting these risks.
  • Ingestion points: The skill instructions (SKILL.md) define the ingestion of 'human expectation', 'agent implementation', and 'verification evidence'.
  • Boundary markers: The framework utilizes structured evidence blocks, YAML schemas (e.g., HUMAN_OVERSIGHT_CASE.yaml), and specific attestation fields to separate human-origin content from agent-generated data.
  • Capability inventory: The skill does not possess capabilities for subprocess execution, network operations, or privileged file system access.
  • Sanitization: The skill mandates an 'Adversarial assumption' policy, instructing the agent to treat all agent-produced content as untrusted and perform negative tests for forged authorship and evidence relabeling.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 13, 2026, 06:07 AM
Security Audit — agent-trust-hub — agile-v-human-oversight