audit-noise

Pass

Audited by Gen Agent Trust Hub on Jul 30, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill processes untrusted data from markdown files within the repository, creating a surface for indirect prompt injection where malicious instructions in a target file could influence the agent's behavior during the audit process.
  • Ingestion points: The scripts/detect.sh script reads the content of arbitrary .md files and the .claude/topic-docs.yaml configuration file within the repository.
  • Boundary markers: The skill respects <!-- markdown-discipline-ignore --> tags to skip content, but does not use explicit boundary markers or delimiters when presenting excerpts to the agent in the final output table.
  • Capability inventory: The skill is strictly read-only and is restricted to shell tools (bash, git, grep, awk, find). It lacks capabilities for network access, file writing, or credential access.
  • Sanitization: Findings include excerpts that are limited to 120 characters by the audit_noise_trim_excerpt function in scripts/lib/noise-shapes.sh.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 30, 2026, 07:55 PM
Security Audit — agent-trust-hub — audit-noise