session-evidence

Pass

Audited by Gen Agent Trust Hub on Jun 26, 2026

Risk Level: SAFEDATA_EXFILTRATIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [DATA_EXFILTRATION]: The skill explicitly accesses sensitive local database files to extract session history, specifically targeting ~/.codex/state_5.sqlite and ~/.codex/sqlite/state_5.sqlite. This allows the agent to read historical transcripts and artifacts which may contain secrets or sensitive data from previous interactions.
  • [COMMAND_EXECUTION]: Scripts provided in references/audit-examples.md utilize node -e to execute inline JavaScript for parsing JSON output from the sessions analyze command. While the logic is functional (calculating result lengths), it involves spawning a Node.js subprocess to evaluate code at runtime.
  • [PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by ingesting untrusted transcript data from previous agent sessions.
  • Ingestion points: Transcripts are pulled from local history via the sessions analyze command (SKILL.md).
  • Boundary markers: Absent. The skill instructions suggest summarizing snippets to maintain privacy but provide no technical delimiters to prevent the agent from obeying instructions found within the logs.
  • Capability inventory: The skill uses the sessions CLI toolset for local data access and analysis.
  • Sanitization: None. The content is parsed and reported as matches without filtering for embedded instructions.
Audit Metadata
Risk Level
SAFE
Analyzed
Jun 26, 2026, 10:13 PM
Security Audit — agent-trust-hub — session-evidence