multi-agent-orchestrator
Pass
Audited by Gen Agent Trust Hub on Nov 21, 2026
Risk Level: SAFEDATA_EXFILTRATIONCOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DATA_EXPOSURE]: The
intent_guardian.pyscript accesses local sensitive files, including~/.env,~/.config/intent-guardian/config.json, and~/.local/state/intent-guardian, to retrieve configuration and maintain session state. It also reads agent transcripts from~/.codex/sessionsand~/.claude/directories to monitor activity. - [COMMAND_EXECUTION]: The skill invokes local command-line tools such as the 1Password CLI (
op) and thecodexCLI viasubprocess.runto retrieve secrets and manage agent environments. - [DATA_EXFILTRATION]: The skill transmits summarized task evidence, status updates, and session metadata to external endpoints, including OpenRouter (
openrouter.ai) and optional user-configured feed URLs. While it implements a redaction function to filter API keys and tokens before transmission, it still exports high-level session data. - [INDIRECT_PROMPT_INJECTION]: The 'Intent Guardian' processes transcripts from external agent sessions which may contain untrusted data from user prompts and tool outputs.
- Ingestion points:
scripts/intent_guardian.pyreads JSONL transcript files from various session storage locations. - Boundary markers: The script uses JSON parsing to isolate human intent from assistant responses and uses specific logic to identify runtime feedback hooks.
- Capability inventory: Executes local CLI tools (
codex,op) and performs network requests to LLM providers. - Sanitization: Implements regex-based redaction to remove credential patterns (e.g.,
sk-,ghp_) from data before it is sent to the observer model.
Audit Metadata