cognitive-toolkit

Pass

Audited by Gen Agent Trust Hub on Jul 15, 2026

Risk Level: SAFEPROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • [PROMPT_INJECTION]: Indirect prompt injection surface detected. The skill ingests untrusted user input (automatic thoughts and situation descriptions) and processes it through a guided workflow without sanitization or boundary markers. These inputs are logged to the filesystem for longitudinal review, creating a potential path for persistent injection if the agent parses previous logs in later sessions.
  • Ingestion points: User conversational input during CBT/DBT sessions and local genomic files via references/health-integration.md.
  • Boundary markers: Absent; user-provided text is interpolated directly into prompts and session transcripts.
  • Capability inventory: File writing (~/Brains/brain/cognitive-toolkit/sessions/), file reading (~/Brains/genome/), and tool execution (mcp__health__*).
  • Sanitization: Absent; user responses are recorded verbatim into session markdown files.
  • [EXTERNAL_DOWNLOADS]: The skill utilizes external Model Context Protocol (MCP) tools (mcp__health__*) to fetch physiological vitals and activity summaries from Apple Health and Aurora Ring sources.
  • [COMMAND_EXECUTION]: Automated filesystem operations are performed to write structured session logs and therapeutic pattern analysis to hardcoded local directories (~/Brains/brain/cognitive-toolkit/sessions/ and cognitive-toolkit/patterns/).
  • [DATA_EXFILTRATION]: Accesses sensitive personal data, specifically pharmacogenomic markers (e.g., COMT, SLC6A4 variants), by reading local files from a designated 'Genome Vault' at ~/Brains/genome/. While processed locally for therapeutic context, this represents access to private health information.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 15, 2026, 07:30 AM
Security Audit — agent-trust-hub — cognitive-toolkit