# JC PenTest Harness

Evidence-first, human-in-the-loop web cockpit for the Jean-Claude multi-model pentest harness.

Production URL: https://pentest.platphormnews.com
Updated: 2026-07-20T00:00:00.000Z

## Responsible Use
- Authorized testing only.
- Only test systems you own or have explicit written permission to assess.
- The operator is responsible for target authorization and scope.
- No anonymous active testing against arbitrary targets.
- No denial-of-service testing unless explicitly scoped and separately enabled.
- No destructive testing.
- No persistence, malware, credential theft, exfiltration, or public exploitation.
- No social engineering.
- No attacks against third-party infrastructure outside the approved scope.
- No remediation, PR creation, or branch push actions.
- Findings are assessment outputs, not automatic fixes.
- Human review is required before real execution and before publishing reports.
- Private findings and artifacts remain protected.

## Run Boundary
- Dry Run is the default and validates scope, config, model routing, prompts, and expected runner command without active target calls.
- Real Run requires PLATPHORM_API_KEY or authenticated operator context, accepted responsible-use policy, allowed hosts, explicit scope, and typed AUTHORIZED confirmation.
- Real Run starts bounded passive recon plus AI Gateway Hunt/Validate execution when Gateway credentials are present; otherwise it records an operator checkpoint instead of fake success.

## Discovery
- OpenAPI: https://pentest.platphormnews.com/openapi.yaml
- MCP: https://pentest.platphormnews.com/api/mcp
- Trust policy: https://pentest.platphormnews.com/.well-known/trust.json

## Models
- fable5: Fable 5 (anthropic, Primary Hunt + Validate)
- opus48: Claude Opus 4.8 (anthropic, Scorer of record + gapfill)
- gpt55: GPT-5.5 (openai, Cross-validation)

## Known Gaps
- Consolidated findings.jsonl is emitted by the CLI at both the per-model and top-level run directories.
- scoring.js implements priority scoring; web real-runs also derive scorecards from actual persisted findings.
- llm-router.route() exists, but the CLI runner scaffolds prompts and pauses for supervised invocation.
- Trust-boundary checks are preflight warnings for repo config and must be enforced by operator review and sandbox/job policy for active extensions.
- Sandbox and BrowserOps previews are public-safe. Protected handoffs only report success after a downstream response confirms receipt.