Eclipse Redteam MCP
Adversarial simulation for AI agents — sandboxed twins only.
status: in-developmentclass: mcpTrust & Receipts: Oversight, Forensics & Insurance Readiness
Vendor: Cyan Eclipse Technologies
What it does
- Adversarial simulation against a sandboxed twin of an agent — never the production agent — covering prompt injection, mandate escape, simulated wallet drain, card-confusion, and recorder-reach probes
- Authorization is enforced as surface absence plus a signed record: no tool accepts a free-form URL, host, or endpoint, and an unregistered or expired authorization refuses before any probe is constructed
- Outcomes are measured per scenario (resisted / escaped-at-step-N / not-applicable), never rendered as verdicts about a third party; reports are signed and verify offline with the public key only
- Simulated funds stay simulated on a hash-chained ledger with value class marked on every entry; probe traffic itself is receipted
- v0.1 is local, fixtures only, with three fixture twins and five fixture suites; live twins, hosted arenas, and external runs are not built
Tools
rt_register_target · rt_suites_list · rt_run_suite · rt_report
Platforms
- MCP over stdio (HTTP transport on roadmap)
Certification
Bench verdict: CERTIFIED(local)
Certified scope: CERTIFIED(local) — Eclipse Bazaar Test Bench 2026-10-26, 26 pass, 1 na of 27 checks (U1 auth: stdio process boundary). Run by W-003 (Verity), suites/redteam.py. Run 3 is the standing card; runs 1–2 FAILED cards are retained append-only. N-CHECK: NOVELTY_PROVEN — probe construction is downstream of a signed, in-date authorization record plus arena or harness registration proof; with neither, no probe object is constructed at all.
Not tested:
- External runs — none performed or built for. First external run requires counsel review of the authorization instrument (spec §9.6) and a named owner decision on services posture.
- Live twins / hosted arenas — NOT built. v0.1 loads in-repo fixture twin config artifacts into the local harness; fixture transports only; no network egress exists in the package.
- Continuous/recurring simulation subscriptions — gated post-v0.1 (spec §9.1).
- Benchmark publication / cross-customer data use — absent by cluster law §2.6; no cross-customer data flow of any kind in v0.1.
- Production signing-key custody — v0.1 signs with a local seed key (keys/, gitignored, chmod 600); custody is a gated decision.
- Hosted/multi-tenant operation and auth beyond the stdio process boundary.
Bench scorecard (evidence): https://eclipsebazaar.com/cards/eclipse-redteam-mcp-2026-10-26-3.json
Pricing
PRICED — $2.00/run (PC-8) via the Eclipse Bazaar bench-order path (order intake) only. No checkout, payment, or buy-now link exists for this product. Free local v0.1 (fixture twins) remains free; external runs are counsel-gated — intake orders are reviewed and confirmed by the operator before any payment link is sent.
Install / verify
docker run -i ghcr.io/cyan-eclipse-technologies-llc/eclipse-redteam-mcp:latest
Verify any MCP server in your client configs against this catalog with bazaar check (Eclipse Bazaar MCP / CLI — see llms.txt).