capability-evolver

Warn

Audited by Gen Agent Trust Hub on May 18, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATIONREMOTE_CODE_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: Extensively invokes shell commands via execSync for system inspection (ps, pgrep, df), hardware discovery (ioreg, MAC addresses), and development workflows (git, node, npm).
  • [EXTERNAL_DOWNLOADS]: Communicates with evomap.ai to fetch evolution tasks, genes, and reusable capsules. It also supports downloading full skills from the EvoMap network via the fetch command.
  • [DATA_EXFILTRATION]: Transmits a structured environment fingerprint to the vendor's hub at evomap.ai, including hashed hostnames, hashed working directories, hardware identifiers (derived from /etc/machine-id or MAC addresses), and OS details. It also shares desensitized execution traces containing file metrics and error signatures.
  • [REMOTE_CODE_EXECUTION]: Executes validation commands defined in Gene assets fetched from the network during the solidify phase. Although these assets are staged and require a manual --validated flag for promotion, the engine is designed to execute arbitrary code provided by these assets.
  • [METADATA_POISONING]: The SKILL.md file contains capability declarations that do not align with actual implementation behavior. It claims to deny several commands (e.g., ps, pgrep, df) and network endpoints (e.g., evomap.ai) that the source code explicitly utilizes.
  • [DYNAMIC_EXECUTION]: The idleScheduler.js module dynamically generates and executes a temporary PowerShell script on Windows platforms to detect system idle time.
Audit Metadata
Risk Level
MEDIUM
Analyzed
May 18, 2026, 01:02 AM
Security Audit — agent-trust-hub — capability-evolver