Skip to content
bluehash

Security

Security & threat model.

A precise statement of what BlueHash defends against, what assumptions it relies on, where those assumptions break, and how we behave when they do. Public, versioned, and reviewable.

01 · Trust model

Trust model

BlueHash is not a custodian. The customer holds every key, every ciphertext, and every trust root. Compromising BlueHash — the company, our infrastructure, or our personnel — does not produce access to customer plaintext, nor the ability to mint authority on a customer’s sigchain.

Trust roots are pinned by the customer at deployment and are rotatable on chain. Vendor-neutrality is enforced by giving the customer the explicit list of EK certificate authorities, enclave platform certificates, IdP issuers, and release signing keys their environment will accept.

02 · Threat model

Threat model

We model state-aligned and financially motivated adversaries with credential-replay capability — including infostealer-driven mass campaigns, IT-helpdesk vishing operators, and Storm-0558-class signing-key thieves. The table below is the canonical list of adversary classes the protocol intends to defeat or to mitigate, with explicit notes on where mitigation is partial or out of scope.

  • Infostealer-driven mass campaigns

    RedLine · Vidar · Lumma · Raccoon

    Stolen browser cookies, password vaults, and OAuth tokens replayed at scale from rented infrastructure.

    Eliminated

  • Vishing → IT helpdesk

    Scattered Spider · 0ktapus

    Operators impersonate employees, persuade helpdesk to reset MFA, replay reset SSO sessions.

    Eliminated

  • Forged SSO tokens

    Storm-0558 · Midnight Blizzard

    Stolen or minted IdP signing keys used to forge tokens that target tenants accept as legitimate.

    Eliminated (attestation binds the consumer, not the issuer)

  • Insufficient-MFA SaaS replay

    Snowflake customers · Citrix portals

    Customer credentials replayed against tenants where MFA was optional or unenforced.

    Eliminated

  • Stolen API tokens / service accounts

    T-Mobile API · Okta support

    Long-lived service tokens leaked to attacker tooling, replayed for months without detection.

    Eliminated (every call carries a fresh quote)

  • Live-session abuse during TTL

    Browser-in-the-middle · proxy phishing

    Adversary uses a real-time relay to hijack an active session before the device-bound TTL expires.

    Reduced (TTL ≤ 5 min) — not eliminated

  • Application-layer compromise

    Server RCE · supply-chain code injection

    An attacker who controls the application server can read plaintext while it is in memory there.

    Out of scope — defense in depth still required

  • Endorsement-key vendor compromise

    TPM vendor signing-key theft

    Adversary forges EK certificates accepted by the customer’s pinned trust roots.

    Out of scope — customer rotates pinned vendor set

03 · In scope & out of scope

In scope & out of scope

Everything BlueHash claims is bounded. The two columns below are the authoritative statement: anything not in the left column is not promised, even by implication.

In scope

7 items

What we eliminate.

  • 01Replay of phished or stolen credentials from a non-attested machine.
  • 02OAuth and SSO token theft via session-cookie exfiltration.
  • 03Long-lived API and service-account tokens used outside their issued device.
  • 04Ciphertext exposure in misconfigured object storage (S3, Drive, SharePoint).
  • 05Unilateral admin compromise via single-key authority.
  • 06Silent equivocation: showing different membership state to different members.
  • 07Backdoor key escrow by a vendor or internal operator.

Out of scope

7 items

What we don’t solve.

  • 01The phishing call itself. Humans get fooled.
  • 02Live-session compromise during the active TTL window (≤ 5 min by default).
  • 03Application-layer zero-days in the customer’s own software.
  • 04Insider misuse within authorized scope — abuse of legitimate access.
  • 05Endpoint-firmware compromise of customer-pinned hardware-vendor signing keys.
  • 06Side-channel attacks against confidential-compute enclaves not yet covered by vendor mitigations.
  • 07Cryptanalytic breaks of standardised primitives (Ed25519, AES-256-GCM, etc.).

We eliminate the credential-replay step. Other defenses still required.

04 · Cryptographic assumptions

Cryptographic assumptions

Every primitive in the suite has a security assumption and a documented fallback if that assumption breaks. The fallbacks are not theoretical — they describe the concrete protocol response, including version cut-over and on-chain coordination.

  • Ed25519 / X25519

    Discrete log on Curve25519 is hard. SUF-CMA security against adaptive chosen-message adversaries.

    Hybrid PQ signatures gated for v2.0 cut-over.

  • AES-256-GCM

    AEAD security against IND-CCA2 with random 96-bit nonces and < 2³² messages per key.

    Per-file content keys cap nonce reuse risk; ChaCha20-Poly1305 alternative available.

  • BLAKE3

    Collision and second-preimage resistance at 128-bit security.

    Sigchain entries can be re-anchored under SHA3-256 if a structural weakness emerges.

  • Hardware attestation roots

    TPM EK certs, Apple Secure Enclave roots, and confidential-compute platform certs are not silently signed-over by the vendor.

    Customer pins the acceptable vendor set; rotation policy is observable on chain.

  • Public transparency log

    Sigstore Rekor (or customer-pinned alternative) does not collude to hide entries.

    Multi-anchor mode writes to N independent logs; verifiers require all to agree.

05 · Failure modes

Failure modes

What happens when something breaks: device lost, share leaked, vendor flaw, transparency-log outage, customer key loss. The protocol’s usefulness is partly measured by whether these failures are recoverable, observable, and survivable.

Device loss

Reported lost device triggers a sigchain revoke entry. Subsequent attestation quotes from that device are rejected by every peer that sees the revoke.

Threshold share compromise

Single FROST share is insufficient to act. Detected share compromise triggers a key-resharing operation, recorded on chain.

TPM firmware vulnerability

Vendor-disclosed flaw triggers EK certificate-chain rotation. Quotes from devices on the unpatched chain are rejected at the verifier.

Transparency log outage

Anchor unavailability stops new admin operations; read access to existing data is unaffected.

Customer key loss

Recovery is via the offline FROST share (paper-printable). BlueHash cannot recover keys — this is a feature, not a defect.

06 · Compliance mapping

Compliance mapping

The protocol is designed to slot under the existing access-control and cryptography controls of common compliance frameworks. We publish concrete mappings rather than blanket claims of coverage.

  • SOC 2 Type II

    CC6, CC7

    In scope · target Q1

  • ISO/IEC 27001:2022

    A.5, A.8, A.10

    In scope · target Q2

  • NIST 800-53 Rev. 5

    AC, IA, SC controls

    Mappings published

  • FedRAMP Moderate

    Agent + verifier path

    Roadmap

  • GDPR / DPA

    Customer holds plaintext

    Sub-processor scope minimal

  • HIPAA / PHI

    BAAs available

    Adapter-level guidance published

07 · Coordinated disclosure

Coordinated disclosure

Security researchers are partners. We respond fast, credit reporters by default, and run a paid bounty for in-scope findings. Coordinated disclosure with reasonable fix windows is preferred to public-then-fix.

Reach security@bluehash.com. Encrypt with the PGP key linked from /.well-known/security.txt.

Disclosure templatesecurity.txt
Subject:  [BH-Disclosure] <one-line summary>
PGP key:  https://bluehash.com/.well-known/security-pgp.asc

Required:
  - Affected component (agent / SDK / verifier / adapter)
  - Reproduction or proof-of-concept
  - Impact assessment, including blast radius
  - Reporter contact + handle preference (credit / anonymous)

Response:
  - Acknowledgement within 24h.
  - Initial triage within 72h.
  - Coordinated disclosure window: 90 days
    (extendable by mutual agreement; reduced for in-the-wild exploitation).

Bounty:
  - In-scope findings: USD 500 – 50,000 by severity.
  - See /.well-known/security.txt for current scope and out-of-scope list.