Enigma
Security

Security & threat model

What is in and out of the boundary.

Enigma · 2026 · open source · enigmamemory.com

This source-tree review document describes the security posture Enigma can review today from the source tree and release evidence. It is not a compliance certification, provider-deletion proof, model-forgetting proof, or tamper-proof hardware guarantee.

Scope and claim boundary

In scope:

  • local vault, receipts, proof bundles, verifier, passport, and context packs;
  • CLI, MCP server, connectors, importer/capsule flows, relay/witness, gateway, mesh-facing relay records, browser extension, desktop scaffold, Docker local demo, hardware collateral, and hosted/BYOC/on-prem deployment responsibilities;
  • public documentation and release evidence that must avoid raw memory plaintext in proof examples.

Out of scope:

  • closed-provider internals, provider-native memory deletion, hidden personalization, logs, backups, caches, embeddings, model weights, and semantic forgetting;
  • third-party accounts, browser/provider pages, cloud accounts, package registries, customer networks, and hardware environments unless the owner has provided written authorization;
  • compliance status or certification language unless separately approved by legal/security with scoped evidence.

Assets

Asset Security goal Notes
Local vault bundle Confidentiality, integrity, user custody, recoverability Canonical Enigma memory state for local mode. Raw memory plaintext must not become public proof material.
Private signing keys and trust material Integrity, non-repudiation, rotation readiness Used for receipts, witness checkpoints, gateway decisions, and verification. Private keys must not appear in packages, images, docs, logs, or support artifacts.
Receipts and proof bundles Offline verifiability, chain integrity, plaintext minimization Public/exported views should carry commitments, roots, hashes, sequence links, signer metadata, and signatures rather than raw memory bodies.
Context packs Purpose-scoped retrieval, receipt-bound injection, local confidentiality Private context may be insertable after user/operator approval; public verification output must be redacted/minimized.
MCP messages and tool calls Protocol correctness, fail-closed validation, client compatibility MCP inputs should be schema/protocol validated and must not make provider-native memory canonical.
Connector configs and backups Safe client setup, atomic writes, minimal secrets Connector writes should be idempotent and preserve user config without leaking memory content.
Import reports and capsules Source caveat preservation, custody continuity, failed-import safety Imports are candidates until written through an Enigma vault and receipted. Source completeness limitations remain visible.
Relay records and witness checkpoints Opaque transport, ordering/accountability, plaintext rejection Relay/witness artifacts must not receive raw memory plaintext. Witnesses sign roots/checkpoint metadata, not memory bodies.
Gateway policies and decisions Default-deny policy enforcement, auditability, tenant control Decisions should reference policy hashes and minimized metadata. SIEM export must not carry raw memory bodies.
Browser extension/native bridge User-approved insertion, origin/path restrictions, no sync plaintext Extension records only target metadata, insertion metadata, and receipt metadata.
Desktop scaffold Accurate operator UX, no false cryptographic proof from UI state UI state is operational evidence only; verifier/proof output remains the cryptographic evidence source.
Docker/local services Reproducible local demo without baked secrets Images/manifests must not embed vaults, keys, credentials, or raw memory examples.
Hosted/BYOC/on-prem deployments Clear responsibility boundary, operator evidence, incident readiness Live production language requires credentials, TLS, durable storage, KMS/secrets, monitoring, backups, support, and incident ownership evidence.
Hardware concepts Key visibility/locality/custody demos Hardware can anchor device identity or signing workflows; it is not claimed tamper-proof.

Trust boundaries

  1. User/local machine boundary: vault files, private context, native host, CLI, verifier, desktop, browser extension, and connector configs share a local trust domain controlled by the user or enterprise workstation policy.
  2. Public proof boundary: exported receipts, public verifier output, public roots, documentation examples, launch site artifacts, and support-safe evidence must be plaintext-minimized.
  3. MCP/client boundary: MCP clients can request memory operations, but server validation and receipt generation remain Enigma responsibilities. Provider-native memory is cache only.
  4. Browser/provider boundary: the extension runs inside provider pages but must require user approval before insertion and cannot prove provider-side deletion or model behavior.
  5. Relay/witness boundary: network services handle opaque encrypted records, roots, checkpoint metadata, signatures, pairing, and authorization, not raw memory plaintext.
  6. Gateway/enterprise policy boundary: gateway policy evaluation produces signed decisions and minimized SIEM/eDiscovery metadata. Customer/operator policy and KMS/SIEM controls are outside the package by default.
  7. Importer/capsule boundary: source exports enter as untrusted inputs with limitations and completeness flags. Canonicalization occurs only after a vault write and receipt.
  8. Hosted/BYOC/on-prem boundary: hosted mode is operator controlled; BYOC/on-prem mode is customer controlled. Responsibility for KMS/BYOK, cloud accounts, logs, backups, data residency, and incident response must be explicit.
  9. Hardware/physical boundary: devices may hold/display signing or proof state. Physical compromise, side-channel resistance, secure manufacturing, and tamper resistance require separate evidence.
  10. Docker/package boundary: package and image artifacts are distribution boundaries and must not contain local vaults, secrets, private keys, or customer data.

Adversaries and abuse cases

Adversary / abuse case Example Primary controls
Malicious local process Reads vault files, native-host messages, connector configs, or private context. OS permissions, local custody guidance, secret handling, minimal public exports, no connector memory backups.
Malicious or compromised MCP client Sends malformed JSON-RPC, unsupported protocol versions, unknown tools, bad params, or destructive requests. Protocol/id/param/tool validation, fail-closed errors, receipt verification, bundle-bound verification.
Provider page or extension abuse Auto-inserts context, captures selected text unexpectedly, records raw context in extension storage. Explicit user click, HTTPS/known-provider checks, no sync storage, transient context, receipt-shape rejection, plain-text insertion only.
Import poisoning Imports incomplete, forged, misleading, or overbroad provider exports/capsules. Source limitation fields, completeness flags, manifest/payload/trust verification, failed-import no-write behavior, canonical vault receipts only after import.
Relay plaintext leakage Client pushes payload fields such as memory bodies, prompts, transcripts, or content. Plaintext-looking field rejection, opaque/encrypted record model, hash-only public relay records, paired/signed auth by default.
Witness equivocation or bad checkpoint Witness signs wrong roots, discontinuous checkpoints, or memory-bearing payloads. Root/continuity checks, signed checkpoints, metadata-only witness artifacts, offline verification.
Gateway policy bypass Unknown provider/region/purpose is allowed or SIEM event leaks memory content. Default-deny unknowns, policy hashes, signed decisions, trusted key requirement for offline verification, plaintext-minimized SIEM export.
Receipt/proof tampering Receipt is changed, reordered, truncated, signed by wrong key, or extended with unsigned/plaintext fields. Chain verification, signer verification, prefix-before-current receipt-log semantics, strict public receipt verification, extension-field rejection.
Context-pack overexposure Public verifier or logs reveal private context text. Redacted public verification, private context used only for local insertion, public-key/trust requirement before valid:true public checks.
Connector config corruption Repeated install corrupts client config or leaks backups. Semantic idempotent writes, atomic writes, backups only on real changes, command/bundle path explicitness.
Docker/image secret bake-in Image includes vault bundle, keys, tokens, or demo memory. Source-checkout Docker demo only, deployment-runbook secret restrictions, external secret/KMS requirements for production.
Hosted/BYOC operator mistake Missing TLS, weak KMS, overbroad operator access, no backup, no SIEM route, wrong data region. Production readiness gates, BYOC runbook, responsibility matrix, named owners, least privilege, monitoring, backup/restore, incident drill.
Hardware loss or compromise Device signing key is extracted or appliance is modified. Key rotation/incident handling, visible custody limitations, no tamper-proof claim, separate hardware evidence required.
Documentation overclaim Public copy implies compliance, provider deletion, model forgetting, token ROI, or tamper-proof hardware. Claim-boundary review, production readiness gates, explicit non-claims in README/security docs.

Component controls and residual risks

Local vault and receipts

Controls:

  • Enigma vault state is the canonical local source for Enigma-controlled memory.
  • Receipt verification covers event signatures, signer identity, active/log roots, sequence links, and tamper scenarios such as changed, reordered, inserted, stale, or wrong-signer receipts.
  • Export/public surfaces are expected to avoid enumerable plaintext and sanitize plaintext-like keys.

Residual risks:

  • Local OS compromise, weak filesystem permissions, user clipboard/history, backups, malware, or support mishandling can expose raw memory outside Enigma's proof controls.
  • Receipts prove custody/lifecycle events, not factual truth of memory statements.
  • Enigma receipts do not prove deletion from third-party providers, backups, logs, screenshots, or model state.

Verification evidence:

  • docs/release-evidence-2026-06-23.md records strict public receipt verification, vault plaintext minimization, receipt-log semantics, and bundle-bound verification evidence.
  • docs/production-readiness.md Gates 1 and 2 enumerate proof integrity and memory lifecycle evidence required before production language.

Context packs

Controls:

  • Context packs are purpose-scoped retrieval artifacts tied to Enigma receipts.
  • Public context-pack verification is redacted/minimized and requires explicit public-key/trust material before returning verified success.
  • Raw memory text belongs in approved local/private insertion paths, not public proof examples.

Residual risks:

  • User-approved insertion intentionally sends selected context to a provider page; after insertion, provider behavior is outside Enigma's control.
  • Overbroad query/purpose policy can expose more local context than intended if operator policy is weak.

Verification evidence:

  • Release evidence records redacted public context-pack verification and private/public output separation.
  • Browser and MCP docs identify user approval and cache-only provider-native memory constraints.

MCP server

Controls:

  • MCP server exposes Enigma operations for init, remember, search, context pack, delete, and receipt verification.
  • Protocol versions, ids, params, tool args, notifications, batches, and error codes are validated.
  • Provider-native memory remains cache only; Enigma vault state remains canonical.

Residual risks:

  • A malicious local MCP client can request authorized operations if it has local access to the configured bundle and process environment.
  • Client UX may display private context after an approved context-pack operation.

Verification evidence:

  • Release evidence records MCP validation coverage.
  • README documents the concrete stdio server command and exposed tools.

Connectors

Controls:

  • Connector profiles target known MCP clients and use explicit command and bundle paths.
  • Writes are semantic, idempotent, backed up only on real changes, and atomic.
  • Connector configs must not become memory-content backups.

Residual risks:

  • Client-specific config locations, environment expansion, and GUI launcher behavior can differ by OS.
  • Existing user config may contain unrelated secrets that Enigma must preserve but not expose.

Verification evidence:

  • Release evidence records connector write hardening.
  • README and docs/client-connectors.md document supported profiles and command overrides.

Importer and capsule flows

Controls:

  • Imports from providers/frameworks remain candidates until written into an Enigma vault and receipted.
  • Source limitations, confidence, and completeness flags are preserved.
  • Capsule verification checks manifest, payload, verifier metadata, trust descriptor, limitations roots, and embedded matching receipt objects before vault writes; failed imports write nothing.

Residual risks:

  • Source exports may omit data, include inaccurate content, or encode provider-specific caveats Enigma cannot independently verify.
  • Imported memories can be sensitive even when metadata is minimized.

Verification evidence:

  • Release evidence records importer/capsule verification and failed-import no-write behavior.
  • README import section states canonicalization requires an Enigma vault write and receipts.

Relay and witness

Controls:

  • Relay stores opaque encrypted records and hash-only public relay records.
  • Relay rejects plaintext-looking fields and requires paired/signed authorization by default; unauthenticated relay is explicit local/demo mode only.
  • Witness checkpoints sign roots and checkpoint/order metadata, enforce root and continuity checks, and must not receive raw memory plaintext.

Residual risks:

  • Metadata such as timing, tenant ids, addresses, counts, and service endpoints can still reveal operational patterns.
  • Local/demo unauthenticated mode is not production authorization.
  • Hosted relay/witness operation requires TLS, durable storage, monitoring, backup, KMS/secrets, incident response, and operator ownership.

Verification evidence:

  • Release evidence records relay/mesh plaintext rejection, paired/signed relay authorization, and witness root continuity checks.
  • Deployment and production readiness docs require opaque relay/witness artifacts and production operator evidence.

Gateway and enterprise policy

Controls:

  • Gateway evaluates Enigma enterprise policy and emits signed decisions.
  • Unknown operations default deny.
  • Public decisions and policy responses are plaintext-minimized; internal raw evaluation exposure requires explicit internal state configuration.
  • SIEM/eDiscovery export carries minimized metadata, not raw memory bodies.

Residual risks:

  • Bad tenant policy can authorize the wrong provider, model, region, purpose, retention, or legal-hold behavior.
  • Gateway evidence does not prove provider-side handling after a request leaves Enigma-controlled systems.
  • Customer-controlled SIEM/log pipelines may add data outside Enigma's minimization controls.

Verification evidence:

  • Release evidence records default-deny gateway behavior and trusted-key requirement for offline decision verification.
  • docs/enterprise-byoc-runbook.md and docs/production-readiness.md define BYOC/on-prem acceptance evidence.

Browser extension

Controls:

  • Manifest V3 scaffold uses the configured native messaging host com.enigma.native_host.
  • The extension requires explicit user approval before inserting context.
  • It avoids sync storage for raw memory and records target metadata, timestamps, receipt metadata, and counts only.
  • Unknown or non-HTTPS provider paths fail closed; all insertion is plain text.

Residual risks:

  • Provider pages can read inserted context after the user approves insertion.
  • Browser extension store review, native-host installation, enterprise browser policy, and provider DOM changes are external operational dependencies.

Verification evidence:

  • apps/browser-extension/README.md documents custody, approval, native-host envelope, and insertion-record constraints.
  • Release evidence records browser approval, URL minimization, no sync storage permission, and sanitized native-bridge errors.

Desktop shell

Controls:

  • Desktop scaffold models local operations and labels UI state as operational evidence.
  • Cryptographic proof remains verifier/proof-bundle output, not visual UI state.
  • Raw-looking imports are rejected or redacted in the product surface evidence.

Residual risks:

  • Static scaffold packaging inside a desktop shell still needs OS-specific signing, auto-update, sandboxing, secret storage, and crash-report controls before production desktop claims.

Verification evidence:

  • Release evidence records desktop copy/inspection hardening.
  • README states desktop UI state is operational evidence only.

Docker and package distribution

Controls:

  • Docker Compose is a source-checkout local relay/gateway demo path.
  • Package files list excludes local vaults and deployment credentials.
  • Deployment docs prohibit baking vault bundles, private keys, tenant credentials, or cloud secrets into images.

Residual risks:

  • A production container deployment still needs external secret injection, TLS, network policy, durable storage, monitoring, backups, and image provenance controls.
  • npm pack --dry-run and CI are package evidence only, not cloud deployment evidence.

Verification evidence:

  • README and deployment docs separate source Docker demos from hosted/BYOC readiness.
  • Release evidence states local audit excludes Docker daemon runtime and external deployment credentials.

Hosted, BYOC, and on-prem operation

Controls:

  • Hosted mode requires Enigma/operator responsibility for infrastructure, TLS, durable storage, KMS/secrets, logging, support, backups, monitoring, and incident response.
  • BYOC/on-prem mode requires customer-controlled cloud/network, KMS/BYOK, logs, SIEM, backups, data residency, operator access, and incident response.
  • Production readiness gates require responsibility matrices, acceptance packets, proof boundaries, legal/security review for compliance wording, and owners for blockers.

Residual risks:

  • Until deployment evidence exists, hosted/BYOC/on-prem should not be described as live production.
  • Shared responsibility mistakes can create gaps in backups, key rotation, incident notice, support access, or data residency.

Verification evidence:

  • docs/enterprise-byoc-runbook.md and docs/production-readiness.md define operator acceptance requirements.
  • README states hosted cloud is not included by default and requires deployment credentials and infrastructure.

Hardware

Controls:

  • Hardware collateral frames devices as proof/custody visibility, locality, device identity, and signing/workflow anchors.
  • Security docs explicitly avoid tamper-proof, raw compute superiority, and provider-internal claims.

Residual risks:

  • Physical compromise, manufacturing supply chain, secure-element configuration, key extraction, enclosure tampering, updates, and field servicing require separate hardware security evidence.

Verification evidence:

  • hardware/README.md states the hardware claim boundary and excludes physical-compromise resistance and provider-internal proof.

Verification checklist for reviewers

Before approving release, public launch, hosted availability, or BYOC pilot language, reviewers should confirm:

  • target docs link this threat model and SECURITY.md;
  • no public proof example contains raw memory plaintext;
  • receipts, public verifier output, relay records, witness checkpoints, gateway decisions, SIEM exports, browser insertion records, connector backups, Docker images, and support artifacts are plaintext-minimized;
  • local verifier evidence covers changed/wrong-signer/reordered/inserted/truncated/stale proof failures;
  • MCP fails closed on malformed protocol inputs and unsupported tool calls;
  • relay/witness require production authorization and reject plaintext-looking memory fields;
  • gateway default-denies unknown inputs and signs decisions with trusted verification material;
  • importer/capsule failed verification writes nothing and preserves source limitations;
  • browser extension requires user approval and records no raw memory plaintext in extension storage;
  • desktop copy does not equate UI state with cryptographic proof;
  • Docker and packages contain no vault bundles, keys, credentials, or private memory examples;
  • hosted/BYOC/on-prem language is blocked until credentials, TLS/domain, storage, KMS/secrets, monitoring, backups, support, incident response, and tenant/customer acceptance evidence exist;
  • no copy claims compliance/certification, provider deletion, model forgetting, token ROI/profit/equity/revenue share, tamper-proof hardware, raw compute superiority, or benchmark leadership without scoped evidence.

Residual-risk summary

Enigma reduces ambiguity around Enigma-controlled memory custody by making local vault state, receipts, context packs, relay/witness records, and gateway decisions verifiable and plaintext-minimized. It does not remove the need for host security, browser/provider trust decisions, customer policy review, deployment hardening, key management, backups, incident response, legal/compliance approval, or separate hardware security evidence.