gpt-engineer-mem
Pass
Audited by Gen Agent Trust Hub on Sep 13, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill functions by retrieving prior session data to inform current tasks, which creates a potential surface for indirect prompt injection from stored history. However, the skill provides robust mitigations and instructions for safe handling. \n
- Ingestion points: The skill reads project context and historical observations from
~/.codex/memories/MEMORY.mdand the localclaude-mem.dbSQLite database. \n - Boundary markers: Instructions mandate labeling all recalled claims as
confirmed,stale,contradicted, orunverifiedand use specific handoff packets to isolate verified facts. \n - Capability inventory: The skill performs local file system operations for engineering tasks and executes a bundled Python diagnostic script. \n
- Sanitization: The instructions explicitly require the redaction of sensitive history and secrets before delegating tasks to child agents. \n- [COMMAND_EXECUTION]: The skill uses a bundled Python script,
scripts/memory_preflight.py, for diagnostic purposes. The script is executed locally to check the status of memory workers. \n- [DATA_EXPOSURE]: The diagnostic script accesses local configuration files (e.g.,~/.claude/settings.json) and database metadata to verify service health. These operations are performed using read-only/immutable modes and the results are used locally for status reporting, not exfiltrated to external domains.
Audit Metadata