Security & trust

Verify, don't trust — including us.

This page describes how AugmentEV protects sensitive workloads: where the confidential boundary sits and what runs inside it, what every job proves, what we can and cannot see, how we handle vulnerabilities — and what we deliberately do not publish. It is written in the same spirit as the product: claims you can check, limits stated plainly.

The confidential boundary

What runs inside the seal — and what doesn't.

Confidential-tier jobs execute inside a hardware-attested trusted execution environment (Intel TDX). The environment is sealed from the machine's owner, the operator, and AugmentEV itself: inputs enter encrypted, outputs leave encrypted, and the memory in between is not readable by anyone outside the boundary. The guarantee is rooted in the processor, not in our policy — attestation is anchored in hardware, which is why it can be checked independently of our word.

Inside the boundary

Confidential-tier workload execution, inside the hardware-attested environment. Inputs and outputs isolated from host and operator. Fail-closed: if the environment can't be verified, the job is rejected, not run.

Outside the boundary

Standard-tier and burst GPU execution runs on non-enclave nodes. Those jobs still carry the post-quantum-signed receipt — proof of what ran — but not the enclave guarantee. We say so on the pricing page, not in a footnote.

Straight talk: "confidential" means a specific, named tier with a specific, named boundary — not a marketing adjective applied to everything we run.

Proof on every job

Evidence that doesn't depend on our word.

Every job — every tier — returns a Proof-of-Task-Execution: a tamper-evident record of exactly what ran, signed with ML-DSA-65 (NIST FIPS 204), a post-quantum digital-signature standard. Anyone can verify a receipt independently — in the browser or with one curl call — with no account and nothing installed. Receipts are retained in tamper-evident form for at least six months, aligned with EU AI Act Article 12 record-keeping expectations.

Key custody, stated plainly

Receipt-signing keys are held by the platform today. On the roadmap — and labeled as roadmap everywhere we mention it — is countersignature by an independent custodian's hardware security module and write-once neutral custody, so that no party to a job, including us, can alter, reorder, or drop the record. FIPS 140-3 validated cryptographic modules for the signing path are also on the roadmap.

Remote attestation

Ask the hardware, not us.

For confidential-tier customers and pilots, attestation quotes are available on request: cryptographic evidence, produced by the hardware, of the environment your workload ran in. An attestation proves the genuineness and identity of the execution environment — it is evidence about where your job ran.

Its scope has limits, and we'll state them: attestation does not, by itself, vouch for what your workload's code does inside the environment — that remains your code's responsibility, and ours is the boundary around it. During pilots we walk your security team through verification tooling and what each claim covers.

Patch & vulnerability posture

Current, monitored, and disclosed responsibly.

  • Patch cadence. Firmware and microcode relevant to the confidential-computing stack are kept current with vendor releases. We do not publish version numbers or patch levels (see below) — but we do commit to the practice.
  • CVE monitoring. We track vulnerabilities affecting the confidential-computing stack and assess their impact on attestation validity — a CVE that undermines the boundary is treated as a trust emergency, not a routine patch.
  • Coordinated disclosure. If you believe you've found a vulnerability — in our platform, our attestation path, or our verification tooling — we want to hear from you: security@augmentev.com. Our disclosure policy is published at security.txt.
Trust boundary & data handling

What we can see, and what we can't.

  • Confidential tier: we cannot see into your workload. Inputs and outputs are sealed inside the boundary; the platform routes, meters, and proves — it does not observe content.
  • Every tier: jobs carry hard cost and runtime caps you set, enforced by the platform. Fail-closed means unverified jobs are rejected, never run unverified.
  • What we do hold: account and billing details (processed by Stripe), the signed receipts (≥6-month tamper-evident retention), and operational metadata needed to run jobs. Our Privacy Policy and published Data Processing Agreement describe handling in legal detail.
  • Analytics: cookieless, privacy-friendly measurement on this site; strictly necessary cookies only (see the consent notice).
Deliberate restraint

What we deliberately do not publish.

Some details aid an attacker more than they inform a buyer, so we don't publish them: processor SKUs and steppings, microcode versions, kernel and hypervisor versions, node and environment identifiers, region, datacenter, or provider topology, and patch-level specifics.

This isn't evasiveness — it's the same discipline we ask of your workloads: expose the proof, seal the rest. Substantive assurance questions get substantive answers in the pilot process, under NDA where appropriate.

Straight talk: we lead with proof, not pedigree — and we protect operational detail the way we protect your data.

Compliance posture

Where we are, honestly.

  • Certifications: we are not SOC 2 or ISO 27001 certified today, and we won't imply otherwise. What you can rely on today is per-job, independently verifiable evidence. Dated commitments land on the changelog.
  • Contracts: a published DPA, Terms, and Privacy Policy; AugmentEV, Inc. is a Delaware corporation.
  • Record-keeping: ≥6-month tamper-evident receipt retention, aligned with EU AI Act Article 12 expectations.
  • Ecosystem: member of NVIDIA Inception; listed on the Model Context Protocol registry; patents pending.

Assurance questions?

Verify a receipt yourself — or bring your security review to a pilot conversation.

Open the verifier Start a pilot