The flight recorder for corporate AI.
Every prompt, response, tool call and agent payment your organisation makes is checked against policy, recorded, signed with post-quantum cryptography and anchored on chain. You can prove it to any auditor, years later, without anyone having to trust you, your vendors, or us.
Auditors, regulators and boards now ask three questions about AI.
Most organisations answer them with application logs that an administrator can edit. Candor AI answers them with evidence nobody can rewrite.
Who used AI, on what data, under which rules?
Every record names the identity (SSO, mTLS or key, pseudonymised), the model and provider, data-loss findings, cost, and the exact policy in force.
Can you prove the logs were not changed?
Records are batched into Merkle trees, signed with ML-DSA and anchored on chain. Editing, deleting or reordering a single record breaks verification.
Did the policy exist before the event?
Policies are published on chain by your own multisig before they can govern traffic. Every record cites the policy hash that governed it.
One gateway between your people, apps and agents, and every model they use.
Point existing OpenAI or Anthropic SDKs at the gateway by changing one URL. Nothing else in your stack changes.
Identify and enforce
Entra ID, Okta or any OIDC provider, mTLS or API keys. Role-based model access, budgets, rate limits and tool permissions.
Protect data
Data-loss prevention redacts or blocks cards, IBANs, SSNs, health identifiers and secrets before a provider sees them.
Record and seal
A write-ahead record is stored before the call proceeds. Content is sealed to your offline archive key with post-quantum encryption.
Sign and anchor
Batches are signed with ML-DSA-65 and their 32-byte Merkle root is anchored on Monad. Nothing personal ever goes on chain.
Built for the security review, not the demo.
Enterprise identity
Microsoft Entra ID, Okta, Google, Ping, Keycloak or any OIDC provider, with tenant pinning and role mapping. The identity provider's signing keys are anchored too, so a login can be re-checked years later.
Policy as evidence
Approved models, providers, regions, DLP rules and retention in one policy. Its hash is published on chain before it can govern traffic, and a new endpoint means a new hash.
Data-loss prevention
Cards with Luhn checks, IBANs, SSNs, Medicare IDs, secrets, private keys and custom patterns, caught across every field of a request, including Unicode evasions. Unscannable content is denied by default.
Agent spend guard
An on-chain treasury for AI agents with per-transaction, per-period and lifetime caps, recipient allow-lists and a kill switch. A prompt-injected agent can ask for anything; the contract will not pay it.
Every model and modality
Text, code, embeddings, images, speech and transcription across frontier labs, crypto-native networks and your own models. OpenAI- and Anthropic-compatible, so existing SDKs work by changing one URL. Models & providers →
Privacy by construction
Content is sealed to an archive key you keep offline; the gateway cannot read its own history. Only salted commitments are anchored, so GDPR erasure leaves proof of existence but nothing readable.
Your evidence, your keys, your infrastructure.
Candor AI runs in your cloud or data centre. Your own multisig owns your audit stream. Farnsworth has no role on it, holds none of your keys and runs no servers in your data path.
If you stop subscribing, every record you already have stays verifiable forever. Your evidence is never held hostage.
You hold
- Signing keys (ML-DSA)
- Archive decryption key, offline
- The stream owner multisig
- All prompts and responses
We never see
- Your prompts or outputs
- Your users' identities
- Your provider API keys
- Your audit records
Fail-closed compliance packs.
Enable a pack and the gateway refuses to start on a configuration that would violate it. For example, a HIPAA policy that routes to a provider without a BAA is rejected. Packs are designed to support your controls; they are not a certification.
Security
SOC 2 · ISO 27001 · ISO 42001 · NIST AI RMF
Privacy
GDPR · CCPA · HIPAA
Financial
PCI DSS · GLBA · SEC 17a-4 · FINRA 4511 · DORA
Public sector & AI
CMMC L2 · FedRAMP Moderate · EU AI Act
See it on your own traffic.
We'll walk your security team through the architecture, the threat model and a live verification of a record against Monad.