continuous-learning-v2
Pass
Audited by Gen Agent Trust Hub on Oct 14, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONPERSISTENCE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill captures session history (prompts, tool inputs and outputs) into local observation logs and uses a background sub-agent (Claude Haiku) to synthesize this data into behavioral rules ("instincts"). This creates a vulnerability surface where malicious content from external tool outputs could be captured and processed as instructions for future behavior.
- Ingestion points: Observations are collected in
~/.claude/homunculus/projects/<hash>/observations.jsonl(seehooks/observe.sh). - Boundary markers: The background agent is provided with explicit instructions and delimiters in its prompt (see
agents/observer-loop.sh), though it ultimately processes raw session data. - Capability inventory: The sub-agent is granted permission to use the
ReadandWritetools to create or update instinct files in thehomunculusdirectory. - Sanitization: The
hooks/observe.shscript implements a regex-based secret scrubber (_SECRET_RE) that redacts common API key, token, and password patterns before they are persisted to the observation logs. - [EXTERNAL_DOWNLOADS]: The
instinct-cli.pyutility includes animportcommand that fetches instinct definitions from remote URLs usingurllib.request.urlopen. While this is a user-triggered feature, it allows for the ingestion of behavioral rules from arbitrary external sources. - [COMMAND_EXECUTION]: The skill executes system commands such as
gitfor project context detection and theclaudeCLI to perform the background observation and analysis tasks. - [PERSISTENCE]: The skill implements a persistent background service model using
nohupand PID files (.observer.pid) to maintain a continuous learning loop across different terminal sessions.
Audit Metadata