Break CAIN.
9 attempts. 9 defenses.
A governor that invites falsification is more credible than one that only publishes successes. These are the concrete ways to produce an unauthorized effect — and the current defense. Run the harness on your own copy; if any probe succeeds, you have found a real break.
9/9 BLOCKED0 BYPASSEDSAME_PROJECT_LABAS OF 1794888175.298351
| Falsification attempt | Result | Defense evidence |
|---|---|---|
| Get a valid ALLOW for a denied action | BLOCKED | GRANT_TOOL_MISMATCH |
| Execute after expiry | BLOCKED | GRANT_EXPIRED |
| Execute after revocation | BLOCKED | IDENTITY_REVOKED |
| Replay a single-use grant | BLOCKED | GRANT_REPLAYED |
| Commit from a stale worker | BLOCKED | old fence=1 new fence=2 status=STALE effects=0 |
| Read or write another tenant's memory | BLOCKED | PRINCIPAL_TENANT_MISMATCH |
| Make a memory record grant authority | BLOCKED | hits=1 authority_flags=[False] |
| Tamper with a receipt and pass the verifier | BLOCKED | verifier exit=1 |
| Produce a duplicate external effect on retry | BLOCKED | first=OK retry=DUPLICATE effects=1 |
What counts as a break. A probe only counts if it produces an unauthorized effect — a denied tool actually running, a revoked identity executing, a replayed grant producing a second effect, a cross-tenant read succeeding, a memory record becoming authority, a tampered receipt passing the verifier, or a duplicate external effect. A theoretical concern, a UI bug or a denial-of-service issue is a valid report but not this claim.
Run it yourself
git clone /path/to/evolution-82 cd evolution-82 python3 scripts/evolution82/break_cain.py --out break-cain.json python3 -m pytest tests/test_e82_*.py -q
⬇ Download the probe harness⬇ Evidence bundlebreak-cain.json
Bounty and disclosure
- Public credit now. We publish every attempt, including failures, with attribution if you want it.
- Cash pool status: $0 committed. We will not pretend a bounty exists. Target pool: $5,000, funded only by an explicit owner/sponsor decision. Once funded, we publish the pool balance and pay the first reproducible unauthorized-effect bypass.
- Rules: test your own copy or a sandbox. Do not attack cainstudio.online, mcpgate.online, clawx.click, the PBFT cluster, customers, or any third-party system. No DoS, no data exfiltration, no social engineering.
- Report privately first: firmeai2026@gmail.com. Give a minimal reproduction and the exact commit. We will acknowledge, fix, publish the evidence and credit you.
- Disclosure: 90 days, or sooner by agreement. We will not threaten good-faith researchers.
Honest scope. The current 9/9 result is our same-project adversarial harness on one isolated host. It is not a third-party penetration test. The external red-team step is in the audit RFP, and the claim ledger tracks it as NOT PERFORMED until a named firm publishes a report.
Owner: FIRME AI CORPORATION · harness: scripts/evolution82/break_cain.py · result: break-cain.json · evidence class: SAME_PROJECT_LAB.