Skip to content

Latest commit

 

History

History
102 lines (93 loc) · 4.93 KB

threat-model.md

File metadata and controls

102 lines (93 loc) · 4.93 KB

skrifheim Threat Model

Status: baseline

skrifheim assumes hostile networks, compromised clients, stolen admin credentials, copied disks, poisoned AI workers, malicious plugins, replayed messages, stale replicas, and partial node capture.

Protected Assets

  • canonical facts,
  • world history,
  • policy epochs,
  • crypto epochs,
  • classification labels,
  • signatures,
  • evidence references,
  • manifests,
  • optional publishing-extension public/private separation,
  • AI artifact provenance.

Initial Threats

  • unauthorized reads across clearance boundaries,
  • forged device or workload context supplied to policy evaluation,
  • write-down from high classification to lower outputs,
  • query inference through aggregation or repeated small queries,
  • access-pattern leakage through touched segments, indexes, response size, timing, cache hits, or repeated query shape,
  • projection mixing incompatible compartments,
  • stale vector/search results leaking deleted or unauthorized facts,
  • AI artifact poisoning,
  • manifest rollback,
  • rollback to a self-consistent old database directory containing old manifests, WAL, policies, credentials, key states, deleted facts, missing audit events, or reusable emergency approvals,
  • concurrent database processes advancing the same storage directory,
  • storage-format downgrade or partial migration attacks,
  • WAL frame deletion, duplication, reordering, torn-tail append, or appending after an unvalidated corrupt tail,
  • malformed physical records, block indexes, filters, compression metadata, or offset tables that trigger misreads, over-allocation, or integrity bypass,
  • policy-domain high cardinality causing pathological small-file amplification or compaction debt,
  • WAL or segment corruption,
  • weak entropy, nonce reuse, or test-only randomness entering production crypto paths,
  • unauthorized export,
  • replayed, stale, or overbroad API credentials and sessions,
  • nondeterministic or overpowered policy/law-pack evaluators,
  • plugin capability escape.

Design Responses

  • classify facts, projections, blobs, query results, and AI artifacts,
  • reject unsafe query plans before execution,
  • expose aggregate query proof state only across public API boundaries, not per-label planner decisions,
  • bind subject, device, and workload context at the authenticated API boundary before policy evaluation,
  • keep projections rebuildable,
  • sign commits and manifests,
  • require a non-rollbackable freshness anchor for production profiles; signed manifests and AEAD authenticate a stored state but do not prove it is the newest valid state by themselves,
  • fail closed on active startup when local storage is older than the freshness anchor, and open historical roots only through explicit recovery workflows,
  • admit a production entropy/CSPRNG provider before nonce, key, salt, replay nonce, or random identifier generation is security-relevant,
  • scope keys by compartment and epoch,
  • release keys through scoped key-provider boundaries rather than giving the main database process unrestricted root or tenant key authority,
  • quarantine stale or forked replicas,
  • chain audit records and include audit roots in manifests and freshness anchors so valid individual audit records cannot be silently deleted or reordered,
  • bind query plans to authenticated identity, attestation evidence, snapshot, policy epoch, query digest, purpose, expiry, and replay nonce; public query requests must not provide the labels or result security metadata of stored facts,
  • model and gate access-pattern and query-shape leakage instead of assuming encryption hides which data was touched,
  • reject concurrent writers, stale storage leases, storage-format downgrade attempts, and partial migrations before durable roots are trusted,
  • use a block-structured physical layout with authenticated record/block metadata, durable indexes, protected-root accounting, and domain-local compaction instead of whole-segment reads,
  • repair only recoverable torn WAL tails at authenticated frame boundaries and reject interior corruption or append-after-corruption attempts,
  • test storage ordering with failure injection, subprocess crash cases, and reference-model recovery comparisons,
  • make API credential/session revocation part of authenticated access, not an application-side afterthought,
  • evaluate law and policy packs through deterministic bounded mechanisms with no network, filesystem, random, wall-clock, or process side effects,
  • keep public web projections separate from private worlds,
  • treat AI output as untrusted until reviewed.

Break-glass access is not a current bypass capability. Existing break-glass audit metadata records intent and attested context only. The future access model must be scoped, time-bounded, approval-gated, and auditable before use; it must not silently become global clearance escalation. AI-assisted identity proofing can provide evidence, but must not be the sole authorization authority for emergency access.