Proof and documentation
ULTIMA: Adversarial Workbench (Find, Harden, Gate)
What the ULTIMA purple-team loop looks for, how find, harden, and gate work, the behavior evidence on our own bench, and the honest limits of this walkthrough.
What it is
ULTIMA is the purple-team loop behind Qrypto Federal. Rather than asking you to trust that our cryptographic implementations are sound, we attack them ourselves, on purpose, and then gate the fix so the same weakness cannot silently return. The loop has three stages:
- FIND: ULTIMA attacks our own implementation and looks for a real weakness.
- HARDEN: the defensive team fixes the root cause, not the symptom.
- GATE: an automated regression gate proves the fix and would catch the weakness if it ever reappeared.
A finding is only considered closed when a gate proves the fix and would catch the regression, not when someone asserts it is fixed.
What this proves to you (plain English)
- The hard part of post-quantum security is not picking a NIST algorithm. It is implementing it correctly. The failures that matter in practice live in the implementation.
- We hold ourselves to an adversarial standard before we ship: we try to break our own code, fix what breaks, and then lock the fix behind a test that fails loudly if the weakness returns.
- A security claim is only allowed where the artifact actually provides the property. If a label like "post-quantum" or "forward-secret" does not match what the implementation really does, that mismatch is itself the finding we report, rather than something we paper over.
What it looks for (the categories)
These are the classes of implementation weakness ULTIMA probes. Each is a well-understood, publicly documented failure mode in cryptographic software.
| Category | What it checks | Why it matters |
|---|---|---|
| Timing and side channels | Operations on secret data should not reveal the secret through how long they take or how they touch memory. The implementation is written to be constant-time and side-channel-aware. | A secret that can be read from timing is not secret. |
| Decryption-failure oracles | A bad ciphertext must not be distinguishable from a good one before the authentication check. FIPS 203 implicit rejection is preserved, so a decapsulation returns a pseudorandom secret rather than an error that would leak the decision. | A reject-vs-accept signal is a decryption oracle that can be weaponized. |
| Pairwise consistency and verify-after-sign | Generated key pairs are checked for pairwise consistency and withheld on failure (a FIPS 140-3 style self-test); signatures are verified after signing before they are relied upon. | Catches a bad key or a faulty signature before it is trusted. |
| Trust-story mismatch | A property label is only permitted where the artifact genuinely provides it; the gap between claim and reality is treated as a defect to fix or disclose. | Prevents a false confidence that an attacker can exploit. |
Correctness and behavior evidence
Graded FACT where exercised on our bench (one lab host, 2026-10-23); INFERENCE where it is a reasoned consequence.
- The find-harden-gate loop has been exercised against a battery of adversarial inputs on our own stack. Every malformed, forged, truncated, or replayed input was handled as expected: a legitimate message was accepted, and every tampered or inauthentic case was refused or isolated rather than acted upon. [FACT, our bench; the specific authority and transport mechanics are not published in this client proof]
- Implicit rejection is preserved: a decapsulation does not report a bad ciphertext as an error, so there is no decryption-failure oracle. A mismatch is caught one layer up, at the authentication tag. [FACT, design property verified on our bench]
- Key generation runs a pairwise consistency self-test and withholds the key pair on failure rather than emitting a key that might be inconsistent. [FACT, design property]
- Signatures are verified after signing (verify-after-sign) before they are relied upon. [FACT, design property]
- The implementation is written to be constant-time and side-channel-aware. [FACT that the design goal is enforced; the method, order, and any measurement are not published]
- Each confirmed finding is routed to a fix and then to an automated, no-LLM regression gate so the property cannot silently regress. [FACT that the gate discipline exists; INFERENCE on coverage for any specific future finding]
Measured performance
Not applicable. ULTIMA is a method, not a throughput benchmark. The client page executes no live attack; the live workbench runs only inside the sovereign edge against authorized, in-scope targets, with results graded and disclosure owner-gated.
Client FAQ
1. Does this page attack anything? No. The controls on the public page are intentionally disabled. ULTIMA runs only inside our enclave against our own stack or an explicitly authorized target.
2. Is this offensive security? No. It is defensive and legitimate-use only. We attack our own implementations to harden them; we do not run it against third parties.
3. What makes a finding "closed"? A fix plus an automated regression gate that proves the fix and would catch the weakness if it returned. A fix without a gate is not considered done.
4. Why focus on implementation rather than algorithm choice? Because choosing a NIST algorithm is the easy part. Timing side channels, decryption oracles, missing self-tests, and mismatched security claims are where real systems fail, and those are what ULTIMA targets.
5. Will you show us a finding on our own system? Within an authorized, in-scope engagement, yes. Results are graded fact versus inference, and disclosure is owner-gated.
6. Does it depend on a cloud AI service? No. The workbench and its regression gates run inside the sovereign edge; the gates themselves are deterministic and do not require a language model in the loop.
Qrypto Federal. DEMONSTRATION artifact; defensive, in-scope, owner-gated; no live attack from the client page. US English, ASCII-clean.
See it run