skillclaw-skill-evolution

Warn

Audited by Gen Agent Trust Hub on Apr 12, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill requires the execution of shell scripts for installation (scripts/install_skillclaw.sh and scripts/install_skillclaw_server.sh) and involves global system modifications via npm install -g.
  • [EXTERNAL_DOWNLOADS]: The skill processes code downloaded from remote repositories (git clone) and external package registries (NPM).
  • [DATA_EXFILTRATION]: The framework functions as a local proxy that intercepts OpenAI-compatible API calls, providing it with direct access to user prompts and the OPENAI_API_KEY. This intercepted data is subsequently uploaded to external cloud storage (OSS or S3 buckets).
  • [PROMPT_INJECTION]: The 'skill evolution' mechanism presents a high risk for indirect prompt injection.
  • Ingestion points: Untrusted session data (user input and LLM responses) is captured via the proxy and ingested by the evolution server.
  • Boundary markers: There are no apparent delimiters or instructions to ignore malicious commands embedded in the captured session data during the 'Summarize' and 'Aggregate' stages.
  • Capability inventory: The system generates SKILL.md files that define future agent behavior and synchronizes these across entire agent clusters.
  • Sanitization: The skill lacks explicit sanitization or validation logic to filter out adversarial instructions before they are distilled into reusable skill files.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Apr 12, 2026, 12:47 AM