improve-skill
Warn
Audited by Gen Agent Trust Hub on Oct 9, 2026
Risk Level: MEDIUMDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The script
scripts/extract-session.jsprogrammatically searches for and reads sensitive session history files from the user's home directory for multiple agents, including Claude Code (~/.claude/), Pi (~/.pi/), and Codex (~/.codex/). These files contain a complete record of user interactions, which may include sensitive source code, internal environment details, or credentials inadvertently shared in previous sessions.\n- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process session transcripts, which are untrusted external data sources that could contain malicious instructions designed to influence the agent's behavior during the analysis phase.\n - Ingestion points: The
scripts/extract-session.jsscript reads.jsonlsession logs directly from agent-specific directories.\n - Boundary markers: The prompt templates in
SKILL.mdwrap the transcript in<session_transcript>tags, but these markers do not provide programmatic isolation against instructions embedded within the transcript that might attempt to hijack the agent's task.\n - Capability inventory: The script uses
fs.readFileSyncfor local file access. The workflow inSKILL.mdinstructs the agent to write new or modified skill files to sensitive locations such as~/.codex/skills/,~/.claude/skills/, and~/.pi/agent/skills/.\n - Sanitization: There is no evidence of sanitization, filtering, or instruction-stripping performed on the transcript data before it is presented to the agent for analysis and file generation.
Audit Metadata