Skip to content

Security: exec-software/exec-enigma

SECURITY.md

Security

Current implementation

Enigma has a bounded evidence engine, an embedded Matrix client, and a separate POSIX persistence host. The engine checks scoped grants before examining evidence requests, binds them to an exact entity/session and the audience enigma, and checks operation rights, logical expiry, limits, and live delegation ancestry. Delegation cannot expand scope, rights, expiry, or per-request limits. Revocation invalidates a grant and its descendants on later API calls. There is no administrator bypass.

These are API checks over trusted host assertions. The registry's install method is a host-only fixture for provisioning authority, not enrollment or authentication. An entity label, grant ID, or claimed source does not prove identity. Matrix class fields are currently public: code in the same process can access or change store and registry memory. The host must own these handles and keep hostile code outside the process. This slice does not establish an unforgeable capability or a hostile-process isolation boundary.

There is no authentication service, network listener, encryption, or native Anomalous integration. The direct embedded client remains volatile. The durable host retains evidence and permission state under the commit and recovery contract below. No version is qualified to hold sensitive or production data. See ENI-0001 for evidence policy and ENI-0002 for persistence.

Durable acknowledgements and recovery

Only scripts/store.py publishes durable receipts. The Matrix worker stages a candidate; its reply alone is not a durable acknowledgement. The POSIX host holds an exclusive nonblocking advisory writer lock, writes a private candidate, syncs the file, atomically replaces the canonical checkpoint, and syncs the database directory before returning the receipt. Initial creation also syncs the parent directory. An I/O or worker-protocol failure stops the session without a success receipt; no later request uses its uncertain in-memory state.

The single checkpoint retains evidence, permission definitions, delegation ancestry, revocation flags, nonreused ID counters, and the logical clock floor. Reads and denied operations are also committed when their staged state changes, so authorization cannot advance the floor only in memory. Matrix validates the whole checkpoint before exposing evidence or authority.

Opening a store retires all loaded session grants, retains their rows and prior revocations, installs a fresh random incarnation, and commits that recovery transition before ready. The trusted host must re-admit callers and explicitly install new grants. Old-incarnation requests are denied before their bodies are interpreted; old grant rows are never recycled. Persistent permission definitions do not authenticate a session after restart.

Missing, malformed, truncated, checksum-invalid, or semantically invalid primary state fails closed. Opening never auto-initializes a database, and recovery never uses a backup or orphaned staging file to bypass corrupt canonical state. An older valid checkpoint may predate an acknowledged revocation. See the persistence guide for the request and failure contract.

Guarantees assume a trusted owner and storage ancestry plus a local POSIX filesystem honoring fsync, atomic rename, and advisory locks. Directory and file handles reject unsafe final components; this does not contain a hostile owner replacing trusted ancestry. SHA-256 detects accidental corruption, not forgery or restoration of an older coherent snapshot. Retirement plus a fresh incarnation prevents saved session authority from returning through this recovery path; general rollback detection needs separately trusted monotonic state. Physical power-cut behavior and arbitrary filesystem/device behavior have not been qualified. An unacknowledged effect can be present after restart, so callers must inspect state rather than blindly retrying.

Identity and time

Enigma follows the identity, session, and scoped-authority separation described in Anomalous's identity and authority design and ADR 0040. Those documents define the integration direction; this engine does not implement Anomalous admission, authenticators, native sigils, or trusted task/channel provenance.

The host must establish the current verified entity/session and a fresh, monotonic logical tick for each dispatch. Grant expiry means the supplied tick must be strictly less than the grant's expiry; the registry rejects rollback below its observed clock floor. This is a logical-tick model, not a wall-clock or realtime-expiry guarantee. A client copies its context at connection time, so the host must reconnect with fresh context for subsequent dispatches. Holding a client with an old tick does not establish freshness.

A future native adapter must derive context from actual live admission and installed authority rather than request fields. Body/effect binding, durable replay/idempotency accounting, externally verified session incarnations, and trusted clock integration remain required before external access.

The durable host preserves grant IDs and counters and binds requests to the current incarnation. Direct embedded replacement remains a different path: opening a new in-memory registry or store can reuse numbers. Discard old clients and establish new bindings; standalone numeric IDs or saved contexts are never portable authority. Do not bypass durable recovery with fresh in-memory objects and then claim that old authority has been recovered safely.

Evidence and resource boundaries

  • A store has one access scope and its own event IDs, revisions, capacity, and scan budgets. Denied operations return fixed empty receipts before evidence validation; current authority is required for historical reads too.
  • Successful assertions and withdrawals append events. Withdrawal requires the asserting entity and separate withdrawal authority. Source text is unverified data; acting entity/session come from the host context.
  • Resolution distinguishes support, opposition, disagreement, and absence of evidence. A work-limited scan reports incomplete; an evidence-list cap is reported separately. Neither evidence state nor a source label proves truth.
  • Each store holds at most 128 events. Proposition components and source text are bounded at 256 units under NIL's text-length operation; scope and entity labels are bounded at 128. The authority registry retains at most 256 grants, including revoked and retired rows. Recovery and re-admission can exhaust this capacity; there is no compaction or identifier recycling.
  • Resolve meters inspected log events. Withdrawal and inspection return limit when their required scan exceeds the grant's work limit. Event budgets are not CPU-time quotas: bounded withdrawal checks add work within a scan.
  • These checks do not establish resistance to timing side channels, denial of service, host compromise, or allocation failure. Process recovery is bounded by the checkpoint contract above. Direct host memory access can bypass the API contract.

Remaining security work

Native admission integration, durable replay/idempotency state, protected transport, maintained cryptographic dependencies, key rotation and recovery, externally anchored rollback protection, redacted audit integrity and retention, and backup/restore verification remain planned. Historical evidence is not a tamper-proof audit log. Revocation cannot undo completed effects or recover data already returned.

Do not invent cryptographic algorithms or infer safety from a model's stated purpose. Human and agent requests must follow the same authority path.

Reporting an issue

Do not include credentials, private records, or an exploitable security report in a public issue. Use the repository's Security → Report a vulnerability option. If private reporting is unavailable, open a public issue requesting a private contact, without vulnerability details or sensitive data.

Only synthetic data belongs in public examples and reproductions. Supported versions and response commitments will be documented when a release process exists.

There aren't any published security advisories