new-chat-ready

Warn

Audited by Gen Agent Trust Hub on Aug 29, 2026

Risk Level: MEDIUMDATA_EXFILTRATIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [DATA_EXFILTRATION]: The skill explicitly targets sensitive application log files stored in the user's home directory to recover conversation state. These files contain full interaction histories which are then summarized into new chat prompts or sent to new threads.\n
  • Evidence: references/recovery-from-local-logs.md lists search paths including ~/.codex/archived_sessions/*.jsonl, ~/.codex/history.jsonl, ~/.codex/logs_*.sqlite, and ~/.claude/projects/.\n
  • Risk: While the skill includes instructions to avoid copying secrets, the automated reading of these files exposes all historical user interactions to the current agent context and potentially new sessions via the send_message_to_thread tool.\n- [COMMAND_EXECUTION]: The skill utilizes powerful shell utilities to query and process log data in the home directory, which could be misused if the search anchors are manipulated.\n
  • Evidence: The skill provides specific command patterns for rg (ripgrep), find, and sqlite3 to be executed against system log paths in references/recovery-from-local-logs.md.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes historical conversation data to generate a 'next-chat prompt', potentially carrying over malicious instructions from past logs into new sessions.\n
  • Ingestion points: Local log files located in ~/.codex/ and ~/.claude/ as defined in SKILL.md and references/recovery-from-local-logs.md.\n
  • Boundary markers: The references/new-chat-prompt-template.md uses standard markdown sections but lacks robust delimiters or programmatic escaping to prevent instructions within logs from being interpreted as system instructions in the new session.\n
  • Capability inventory: The skill combines shell execution (rg, find, sqlite3) with thread management tools (create_thread, send_message_to_thread) to handle the recovered content.\n
  • Sanitization: The skill relies on natural language instructions to 'summarize' and 'sanitize' content, which is a soft constraint vulnerable to bypass by adversarial content in the logs.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Aug 29, 2026, 02:04 PM
Security Audit — agent-trust-hub — new-chat-ready