codex-collaboration

Warn

Audited by Gen Agent Trust Hub on Oct 19, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONPRIVILEGE_ESCALATIONDATA_EXFILTRATION
Full Analysis
  • [COMMAND_EXECUTION]: The scripts/codex_session.py script serves as a wrapper for several system-level commands, using subprocess.run() to execute codex CLI operations, git status/log checks, and lsof for process tracking. This execution mechanism allows the agent to interact with the underlying operating system and external tools based on session data.
  • [INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest and process 'rollout' transcripts from previous Codex sessions. These transcripts represent external, untrusted data that could contain malicious instructions designed to influence the behavior of the agent currently reading them.
  • Ingestion points: The iter_events function in scripts/codex_session.py reads JSONL rollout files from ~/.codex/sessions/.
  • Boundary markers: The skill lacks explicit boundary markers or instructions to treat transcript content as untrusted data during the tail or extract operations.
  • Capability inventory: The script possesses capabilities for file system writes (extract, handoff commands) and arbitrary shell execution through codex delegation and git commands.
  • Sanitization: There is no evidence of sanitization or filtering of the transcript text before it is presented to the agent or written to output files.
  • [PRIVILEGE_ESCALATION]: The delegate subcommand in scripts/codex_session.py includes a --full-auto flag. When used, this flag passes the --dangerously-bypass-approvals-and-sandbox argument to the Codex CLI, explicitly disabling safety constraints and sandboxing protections for the execution environment.
  • [DATA_EXFILTRATION]: The owners function in scripts/codex_session.py performs a sensitive scan of the /proc filesystem. It iterates through all active process IDs and their file descriptors (/proc/[pid]/fd) to identify which processes have the rollout files open. This constitutes information gathering of system-wide process activity which is often restricted or considered sensitive metadata.
  • [EXTERNAL_DOWNLOADS]: The script references a specific, hardcoded path for a quota probe: /home/lachlan/ProjectsLFS/AgenticApp/agentic_tools/wechat_gui_agent/scripts/codex_quota_status.py. While this appears to be a local path for the author, the mechanism allows for the execution of an external script file that is not distributed with the skill itself, creating a dependency on unverifiable local environment configurations.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 19, 2026, 06:24 PM