continuous-learning

Pass

Audited by Gen Agent Trust Hub on Mar 7, 2026

Risk Level: SAFEPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted data from the conversation transcript to generate and save new instructions for the agent.
  • Ingestion points: The entire current conversation transcript is scanned for patterns in SKILL.md.
  • Boundary markers: No programmatic boundary markers are used to separate untrusted session content from the extracted logic; the skill relies on a manual user confirmation step (AskUserQuestion) which may be bypassed or subverted by a poisoned transcript.
  • Capability inventory: The skill writes Markdown files with YAML frontmatter to .claude/skills/learned/ and ~/.claude/skills/learned/, which are interpreted as skills in subsequent sessions.
  • Sanitization: No sanitization, escaping, or validation of the extracted content is performed before writing it to the filesystem.
  • [DATA_EXFILTRATION]: The 'Continuous Learning' process may capture sensitive information present in the session history—such as API keys, environment variables, or private code snippets—and save them to disk in the learned/ directory. This creates a risk of local data exposure if the filesystem is shared or the skills are moved between environments.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 7, 2026, 04:21 PM