# CAIN-42 E82 External Security Audit — Request for Proposal

**Issued:** 2026-11-17
**Issuer:** FIRME AI CORPORATION
**Subject:** Independent security review of the CAIN-42 governed execution boundary, evidence fabric and recovery path
**Status:** Open. No auditor has been engaged. No third-party audit has been performed.

---

## 1. Why this RFP exists

CAIN-42 claims to be the **authority, enforcement and evidence layer for vertical AGI**: an agent proposes, CAIN decides, the boundary enforces, the system executes, an independent verifier checks, and evidence remembers. The claim is only credible if an outside party tries to break it and publishes what it finds.

This RFP defines a bounded, lawful, reproducible audit. It is not a certification engagement and it does not ask an auditor to endorse the phrase “governor of vertical AGI”. It asks one question:

> Can an unauthorized action produce a real effect, or can an authorized action be misrepresented in the evidence?

## 2. Assets in scope

| Asset | Location | Notes |
|---|---|---|
| Source snapshot | `evolution-82` commit `9bb134a72b0f7af68d02071adaf30b0cf1e39bab` (or the commit named in the signed bundle) | Read-only clone |
| Evidence bundle | `/proof/bundle/evolution-82-vertical-agi-2026-11-17/CAIN42_E82_EVIDENCE_BUNDLE.zip` | 240 files, 55/55 offline checks, Merkle `63868034…` |
| Standalone verifier | `verify_e82_offline.py` | Stdlib + `cryptography`; no CAIN imports |
| Development replica | Provided on request | Isolated, synthetic tenants only |
| Threat model | `docs/evolution-82/E82_GOVERNANCE_SECURITY_REPORT.md` and this document | |
| Prior lab | `artifacts/e82/security_lab.json` (86/86 same-project) | Not a substitute for this audit |
| Receipt / log spec | E84 `cain/receipt/**` and E82 `audit_chain.py` | Include if the auditor wants the receipt path |

## 3. In scope

1. **Boundary bypass** — produce a real tool effect without a valid, unexpired, unrevoked, single-use ALLOW bound to that exact tool, arguments and tenant.
2. **Grant lifecycle** — forgery, replay, expiry, revocation, cross-process reuse, stale fencing, budget exhaustion.
3. **Child/delegation escalation** — a child agent acquiring authority its parent does not hold.
4. **Cross-tenant isolation** — read or write another tenant's memory, tasks, effects or evidence.
5. **Memory authority** — make a memory record, model output, reputation or skill become authorization.
6. **Evidence integrity** — modify, remove, add, reorder, truncate or re-sign evidence and still pass the verifier; make a receipt disagree with the observed effect.
7. **Recovery** — duplicate effects after a lost acknowledgement, worker kill, lease expiry, host reboot or failover; stale writer commits.
8. **Supply chain and secrets** — credential leakage in logs/evidence; dependency confusion; build provenance gaps.
9. **Denial of service** — only within the isolated replica with a written resource ceiling; no production DoS.

## 4. Out of scope

- Model quality, benchmark scores, “AGI” claims and prompt style.
- Cloud-provider internals, the host kernel, and physical security.
- Hardware attestation / TEE (none is deployed; do not assume one).
- Regulatory certification (SOC 2, ISO 42001, EU AI Act conformity).
- The quarantined PBFT cluster for customer anchoring. It is out of scope until the 2026-11-15 incident is repaired and re-qualified.

## 5. Rules of engagement

- **No production access.** Do not test `cainstudio.online`, `mcpgate.online`, `clawx.click`, the live PBFT cluster, customers or third-party systems.
- **No customer data.** Use synthetic tenants and the provided development replica.
- **No destructive tests outside the replica.** No DoS against shared infrastructure.
- **No secret exfiltration.** If a secret is found, record the location and a redacted proof; do not copy it.
- **Private report first.** Send findings to `firmeai2026@gmail.com` before publication.
- **Disclosure:** 90 days from acknowledgement, or sooner by agreement. We will publish the report or a detailed summary and the remediation.
- **Safe harbour:** We will not pursue good-faith research that follows these rules.

## 6. Access and logistics

- A read-only source snapshot and the signed evidence bundle.
- One SSH account on an isolated development replica, key per auditor, no production network.
- Synthetic tenants; no real credentials.
- A named engineering contact for reproduction questions.
- A written resource ceiling for any load test.
- All access revoked at engagement close.

## 7. Deliverables

1. **Executive summary** — what was tested, what was found, what remains unknown.
2. **Findings** — severity, affected component, exact reproduction, observed vs expected, evidence hash.
3. **Fix verification** — re-test of every fixed finding; unresolved findings stay open.
4. **Coverage statement** — what was not tested, and why.
5. **Public summary** — a version we can publish with the claim ledger.
6. **Attestation** — a signed statement of independence and methodology.

A finding is **critical/high** only if it produces an unauthorized effect or a false verified claim, not merely a code smell.

## 8. Indicative budget and timeline

These are planning ranges, **not offers**:

| Phase | Scope | Indicative range | Time |
|---|---|---|---|
| Focused boundary review | 8 attack classes, 5 auditor-days | $15k–$30k | 2 weeks |
| Full audit + recovery + evidence | 9 classes, replica drills, retest | $45k–$90k | 5–7 weeks |
| Retest only | Verify fixes | $8k–$15k | 1 week |

Payment milestones: 30% kickoff, 40% draft report, 30% final + retest. Travel not required.

## 9. Proposal requirements

- Firm name, lead auditor, relevant agent/MCP/distributed-systems work.
- Methodology and tooling; how you avoid testing only the happy path.
- Independence statement and conflicts.
- Work plan, timeline, named team, price.
- Sample redacted report.
- Data-handling and confidentiality terms.

## 10. Evaluation

1. Ability to test the **execution boundary and evidence path**, not just the web UI.
2. Reproducibility of findings and quality of the sample report.
3. Independence and no conflict with CAIN-42 vendors.
4. Price and timeline.
5. Willingness to publish a summary.

## 11. Submission

Email `firmeai2026@gmail.com` with subject **“E82 audit proposal”**. Questions by the same route. We will publish the selected firm, the final report link and every unresolved finding.

## 12. Honest statement

**No third-party audit has been performed.** The current 86/86 lab result is same-project. The E82 offline verifier is a third-party *verification kit*, not an audit. This RFP exists because that gap is the largest one in the claim ledger.
