It exists because fraud and deepfake detection normally forces a trade-off: to score a signal you must expose the raw biometric or behavioral data, creating the very breach risk you are trying to prevent. AI Fraud Detection removes that trade-off — the agents evaluate encrypted inputs, so detection happens without plaintext ever being exposed.
Related · tier-1 reading. For the proof model behind every AI decision, see Verifiable AI Decisions.
Traditional fraud detection sees your data in plaintext. That's not a feature — it's a vulnerability. H33 runs three native Rust AI agents inside the FHE pipeline, analyzing ciphertext patterns at sub-microsecond speed. Deepfakes, synthetic identities, harvest-now-decrypt-later — detected before the auth response ships.
AI Fraud Detection is a product that catches deepfakes, synthetic identities, credential attacks, and harvest-now-decrypt-later behavior at the authentication boundary — analyzing encrypted data so no plaintext is exposed. It owns the fraud-detection workflow and the authentication-time user outcome. It does not own, redefine, or reimplement the cryptography, evidence, governance, or verification it depends on; it composes them.
The product boundary: AI Fraud Detection owns the authentication-time fraud-detection workflow and outcome. It uses H33 mechanisms for privacy, governance, evidence, monitoring, and verification; it does not redefine or own those mechanisms.
Every fraud detection vendor promises AI. None of them operate on encrypted data. Which means every one of them is a breach waiting to happen. These are the threats that matter.
Not Python. Not an external API. Not a sidecar. Three Rust-native AI agents compiled into the auth binary, running in parallel with FHE verification. They analyze ciphertext patterns — the server never decrypts anything.
Production benchmarks from Graviton4 (c8g.metal-48xl, 192 vCPUs). Every number is from sustained production runs — not cherry-picked burst windows. Verify on /benchmarks →
Traditional fraud detection requires plaintext access to biometric data. That creates a paradox: the system protecting your data needs unrestricted access to it. H33 eliminates the paradox entirely.
Every fraud detection vendor on the market operates on plaintext. They decrypt your users' biometrics, analyze them, and promise to delete the data afterward. You're trusting their word against a breach that hasn't happened yet.
H33's three AI agents operate on data encrypted under BFV fully homomorphic encryption. The biometric template, the verification computation, and the fraud analysis all happen in ciphertext space. A complete breach of H33's infrastructure exposes only encrypted ciphertexts — computationally infeasible to decrypt even with quantum computers, thanks to lattice-based post-quantum security.
Deepfakes fool 97% of liveness detection systems. H33's approach is mathematically immune — because the biometric data is never decrypted. You can't deepfake what you can't see.
Fraud detection is not a separate service. It is embedded in every authentication response. No extra SDK, no webhook, no second API call. The agents run inside the pipeline and their results ship with the attestation.
// 1. Initialize the H33 client const h33 = new H33Client({ apiKey: "h33_pk_..." }); // 2. Capture biometric & encrypt client-side (BFV FHE) const embedding = await h33.biometric.capture("face"); const encrypted = await h33.fhe.encrypt(embedding); // 3. Authenticate — fraud detection included automatically const result = await h33.auth.verify({ userId: "user_abc123", biometric: encrypted, // FHE ciphertext, never plaintext }); // Authentication result result.verified // true result.attestation // Dilithium-signed proof (post-quantum) result.zkProof // ZK-STARK verification hash // Fraud detection flags (included in every response) result.fraud.harvest_detected // false — no HNDL pattern result.fraud.sidechannel_flag // false — no timing anomaly result.fraud.crypto_health // "optimal" — all params valid result.fraud.deepfake_score // 0.003 — genuine biometric result.fraud.risk_level // "low" result.fraud.agent_latencies // { harvest: "0.69µs", sidechannel: "1.14µs", crypto: "0.52µs" } // All fraud analysis performed on encrypted data via FHE // Plaintext NEVER touches the server. Verify →
Every fraud vendor in the market decrypts your data to analyze it. That is the vulnerability. H33 is the only fraud detection system that operates entirely on encrypted data.
| Capability | H33 AI Fraud Detection | Traditional Fraud Systems |
|---|---|---|
| Data access model | Encrypted only — FHE ciphertext analysis | Requires plaintext access |
| Detection latency | 0.52 – 1.14µs per agent | 50 – 500ms typical |
| Deepfake resistance | Cryptographic — ciphertext pattern analysis | Probabilistic — pixel/audio heuristics |
| Post-quantum secure | Yes — lattice FHE + Dilithium + Kyber | No — RSA/ECDSA broken by QC |
| False positive approach | Zero in production — cryptographic validation | Statistical — tunable thresholds |
| Breach exposure | Zero — only FHE ciphertexts stored | Full plaintext biometric data |
| HNDL protection | Active detection + PQ crypto | Not addressed |
| Integration | Embedded in auth response — 1 API call | Separate service — additional latency |
Use AI Fraud Detection when the problem is at the authentication boundary for a single institution: deepfakes, synthetic identities, credential attacks, and harvest-now-decrypt-later behavior, where you need decisions before the auth response ships. Status is honest — benchmark numbers are from v10 production runs on Graviton4 and are reproducible on /benchmarks/; the "zero false positives" figure reflects observed production incidents to date, not a guarantee that none can ever occur.
Three AI agents. Sub-microsecond latency. Zero plaintext exposure. This is what fraud prevention looks like when you stop trusting the server. Free tier includes 10,000 authentications per month.