continuous-learning
Pass
Audited by Gen Agent Trust Hub on Oct 5, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONPERSISTENCEDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted session data from
CLAUDE_TRANSCRIPT_PATH. Attackers could embed malicious instructions in conversation history or tool outputs that the skill identifies as a 'learned pattern'. Evidence Chain: 1. Ingestion:CLAUDE_TRANSCRIPT_PATH. 2. Boundary markers: None. 3. Capability: Writing newSKILL.mdfiles to~/.claude/skills/learned/. 4. Sanitization: None identified. - [PERSISTENCE]: The skill writes generated content to the agent's configuration directory (
~/.claude/skills/learned/). These generated skills persist across sessions, allowing potentially malicious 'learned' instructions to influence the agent permanently. - [DYNAMIC_EXECUTION]: The skill dynamically generates new instruction files (
SKILL.md) based on runtime conversation data. These files are then interpreted as executable skills by the agent environment.
Audit Metadata