A KMS with no keys to steal.
Key management where nothing sits at rest. Decryption capability is derived from attested hardware and a t-of-n quorum at the moment of use, never held as a secret — so there is nothing to exfiltrate, phish, or subpoena. Sealing AI model weights to the GPU is the flagship case.
Keyless by construction
Keys are never stored. Capability is derived from a hardware measurement and a t-of-n quorum at the moment of use, then zeroed. There is no key file to phish, copy, or subpoena — and nothing at rest for an insider to take.
Threshold custody
Custody is split t-of-n (default 3-of-5) across independent parties, dealt in a distributed ceremony with a signed transcript. Any t can derive; t−1 shares reveal mathematically nothing. Any party can revoke.
Derived from attested hardware
Derivation is gated on a verified measurement — Intel TDX, AMD SEV-SNP, NVIDIA Confidential Computing, or TPM2. Wrong hardware, or a tampered stack, and the key never materializes.
Provable destruction
Every access, and every refusal, lands in an append-only, hash-chained log. Key destruction returns a signed receipt — retained ciphertext is permanently unrecoverable, not merely asserted deleted.
Model weights, sealed to the silicon.
Frontier weights are among the most expensive files ever made, and mostly protected like any other blob. Forge seals them to attested GPUs: they decrypt directly into VRAM under Confidential Computing and exist nowhere else in plaintext, so an exfiltrated checkpoint is noise off the approved hardware. You can rent the compute without renting out the model.
- In-VRAM decryption on H100 / H200 / B100 / B200 — plaintext never on disk, in RAM, or on the wire
- Attestation-derived keys: SHA3 measurement of die UUID, VBIOS, driver, and CC flags at load
- Publisher-signed Merkle commitment over every tensor; delta proofs chain fine-tunes to source
- Safe on rented or shared compute — the host sees ciphertext and an attestation handshake
Built on the seven-layer SCSA core. See the platform →