Skip to main content

1Claw Security Overview

Version: 0.2
Date: 2026-08
Classification: Public
Contact: ops@1claw.co

1. Threat Model Summary​

1Claw protects AI agent credentials, signing keys, and operational secrets across their full lifecycle. The threat model assumes:

  • Compromised agent runtime — agents may be prompt-injected or otherwise manipulated
  • Compromised network path — all agent-to-API traffic traverses untrusted networks
  • Insider risk — platform operators should not have unilateral access to customer keys
  • Supply chain attack — container images may be tampered with in transit

Trust Boundaries​

Agent Runtime (untrusted)
    │
    ├── Shroud TEE Proxy (AMD SEV-SNP attested)
    │       ├── LLM inspection + redaction
    │       └── Transaction signing (key unwrapped only inside the TEE)
    │
    └── Vault API (Cloud Run, HSM-backed encryption)
            ├── Envelope encryption (GCP KMS / AWS KMS / Azure Key Vault)
            ├── Policy engine (deny-by-default)
            └── Audit hash chain (tamper-evident)

2. What Is Verifiable Today​

TEE Attestation (Public, No Auth Required)​

Shroud runs on GKE Confidential Nodes with AMD SEV-SNP. Verify the deployment:

curl -s https://shroud.1claw.co/v1/shroud/attestation | jq .

Response fields:

FieldMeaning
attestation_levelnone, identity, confidential, or sev_snp
identity_tokenGCE metadata JWT (verify against Google JWKS)
image_hashPublished container digest
confidential_claimssecboot, hwmodel, instance_confidentiality, sw_name when present
verification.stepsLevel-specific checklist returned by the endpoint

Verification steps (summary):

  1. Decode identity_token as a JWT (do not trust before signature verify)
  2. Fetch Google public keys from https://www.googleapis.com/oauth2/v3/certs
  3. Verify the JWT signature
  4. Confirm aud is https://api.1claw.co (see verification.expected_audience in the response)
  5. Compare image_hash against the published Docker digest (ghcr.io/1clawai/shroud)
  6. At sev_snp level, confirm measurement/digest claims match your expected build

Audit Hash Chain (Org-Scoped, Authenticated)​

Every audit event is linked to the previous via prev_event_id and an HMAC-SHA256 integrity_hash:

curl -s -H "Authorization: Bearer $TOKEN" \
  https://api.1claw.co/v1/audit/verify | jq .

The verify endpoint recomputes HMAC-SHA256 server-side for events in the query window and checks prev_event_id linkage. It returns chain_valid, tampered_events, and unverifiable_events (for pre-2026-08-21 legacy rows).

Chain structure: integrity_hash = HMAC-SHA256(key, JSON.stringify([prev_hash, org_id, actor_type, actor_id, action, metadata, timestamp]))

See Audit hash chain verification for limits (org-scoped window; no offline client recompute without the server key).

Envelope Encryption​

  • Every secret gets a unique Data Encryption Key (AES-256-GCM)
  • DEKs are wrapped by a per-org shared KEK in KMS (HSM for paid tiers)
  • Optional CMEK (Customer-Managed Encryption Keys) for Business/Enterprise
  • Optional vault MPC splits DEK shares across HSM providers

Policy Enforcement​

  • Deny-by-default: agents have zero access until a human creates an explicit policy
  • Scope-bound JWTs: agent tokens carry path scopes derived from policies when agents.scopes is empty
  • Token revocation on policy change: stale permissions are invalidated via agent_active_tokens
  • Control-plane consensus: sensitive admin operations can require multi-party approval (202 + pending_approval_id)
  • Expression policies (v2): tx_conditions.expression evaluated at signing time alongside field matching

3. Architecture Differentiators​

CapabilityDescription
Whole-agent governanceSecrets, LLM traffic, runtimes, memory, and signing under one policy plane
Shroud LLM inspectionRedact secrets from prompts, detect injection, enforce output policies
Deep transaction inspectiondeep_inspect decodes multicall, Safe execTransaction, ERC-4337 inner calls
Consensus composabilityskip_when / require_when on policies prevent automation breakage
Hash-chained auditEvery event cryptographically linked; payload_hash binds pending approvals to actions
Shamir-split key custodyOrg KEK and optional vault DEK shares split across HSM providers (envelope encryption, not threshold signing)

MPC and Shamir key custody (not threshold signing)​

1Claw uses Shamir secret sharing for encryption key custody, not for multi-party transaction signing:

LayerWhat is splitWhere shares liveWhere reconstruction happens
Vault MPC (2of3_multi_hsm, optional per vault)Per-secret DEKGCP KMS, AWS KMS, Azure Key VaultVault API during authorized secret read/write
Org Shamir KEK (Team+ / Business+, migration 203)Organization KEKGCP + AWS (+ optional client share on Business/Enterprise)Shroud TEE via POST .../shamir/reconstruct (501 when Shroud unconfigured)
Treasury / agent signing keysN/A (single HSM envelope)__treasury-keys / __agent-keys vaultsStandard envelope encryption, gated by policies

This is envelope encryption with Shamir-split KEKs/DEKs. It is not Turnkey-style MPC-CMP threshold signing, where multiple parties co-sign a transaction without assembling the full private key.

Turnkey comparison: Turnkey's QuorumOS performs n-of-m signature quorum on private keys. 1Claw's Shamir modes ensure no single cloud HSM provider holds a complete DEK or org KEK, but transaction signatures are produced by a single signing key after policy checks.

4. Compliance Status​

FrameworkStatus
SOC 2 Type IIContact ops@1claw.co for current status
GDPRData export (POST /v1/auth/export-data) and account deletion (DELETE /v1/auth/me) with step-up auth
PCI DSSReference-mode card storage; PAN never persisted for Laso partner cards

5. Contact and Next Steps​

This document is versioned. Check the date above and request the latest version from ops@1claw.co if older than 90 days.

Last updated on