codex-session-controller

Pass

Audited by Gen Agent Trust Hub on Aug 31, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [DATA_EXFILTRATION]: The skill instructions direct the agent to access session history and task records stored on the local filesystem under $CODEX_HOME/sessions (defaulting to ~/.codex/sessions). This access is used to verify task identities and rollout states when standard tool-based reads are incomplete.
  • [INDIRECT_PROMPT_INJECTION]: The skill operates as a routing hub that ingests natural language reports from other workers (AI tasks). This ingestion creates an attack surface where instructions embedded in worker outputs could influence the controller's decision-making.
  • Ingestion points: Data enters the context via the read_thread tool and local filesystem access to session records.
  • Boundary markers: The instructions mention establishing boundaries for safety and permissions, but do not specify technical delimiters for worker content.
  • Capability inventory: The agent has the ability to create new threads, send messages to existing threads, and update thread titles.
  • Sanitization: The skill implements a verification process requiring workers to provide specific evidence (such as tool calls, commit hashes, or test results) before their completion claims are accepted.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 31, 2026, 06:49 PM
Security Audit — agent-trust-hub — codex-session-controller