marketing-os

Pass

Audited by Gen Agent Trust Hub on Aug 28, 2026

Risk Level: SAFENO_CODEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [NO_CODE]: The skill package contains only Markdown and YAML files. No executable scripts, binaries, or command-line tools are included, which significantly limits the attack surface.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external data, such as product documentation and user-provided research, which is a common surface for indirect prompt injection. However, it incorporates strong defensive guidelines to mitigate this risk:
  • Ingestion points: Processes project-specific context files (e.g., .agents/marketing-context.md) and user-supplied business goals.
  • Boundary markers: Implements a strict data labeling system in references/evidence-and-context.md, distinguishing between 'Verified', 'User-provided', 'Analysis', and 'Hypothesis' states.
  • Capability inventory: The skill primarily functions as a router for other specialist skills; it does not contain direct file-write, network exfiltration, or code execution capabilities.
  • Sanitization: Includes an 'Adversarial check' section and a 'Protect the truth boundary' instruction set that forces the agent to validate facts and record counterevidence, preventing the automatic adoption of potentially malicious instructions embedded in research data.
  • [DATA_EXFILTRATION]: No network-enabled tools or commands (such as curl or wget) are present. The skill instructions explicitly forbid connecting accounts or altering live campaigns without authorization.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 28, 2026, 11:55 AM
Security Audit — agent-trust-hub — marketing-os