Agent-readable docs index: /llms.txt. Full docs in one file: /llms-full.txt. Download /docs.zip to grep all markdown files locally.

Known Limitations and Non-Goals

Honesty about security boundaries is essential. Dmint guarantees strict, deterministic protection for the operations it is configured to wrap, under specific architectural assumptions.
This page documents the deliberate non-goals, architectural trade-offs, and known edge cases in the current release.

1. Non-Goals

Dmint is NOT an LLM Input Guardrail

Dmint does not inspect prompts, detect jailbreak phrasing, or judge whether user intent is malicious. It operates at the execution layer: it evaluates whether a tool request matches an authorized capability, binding, and argument schema.

Dmint is NOT an In-Process Sandbox Escape Barrier

If an AI agent runs under the same unisolated operating system user as the Dmint process and has access to a command shell, it can kill the Python runtime, read private keys from disk, or bypass the decorator. Operating system isolation (containers, virtual machines, or separate user accounts) is a required prerequisite.

Dmint is NOT an Enterprise IAM or SSO Directory

Dmint does not manage employee login sessions or organization-wide role hierarchies. It answers one specific question: "May this specific AI agent execute this exact capability with these exact arguments right now?"

2. Storage Limitations and Edge Cases

Database Snapshot Rollback & Monotonic Sidecar Counter

Every record in SQLite or PostgreSQL includes an HMAC-SHA256 integrity tag over the canonical row contents. To defend against an attacker restoring an older snapshot of the database file (a full-file rollback), Dmint also maintains a monotonic store_recovery counter paired with a 0600 .recovery sidecar file protected by cross-process flock locking. If the database counter ever lags behind the sidecar counter, Dmint fails closed with ApprovalRollbackDetectedError (DMT_APPROVAL_ROLLBACK_DETECTED).
  • Defense-in-Depth 1: Short Assertion TTL: Configure a short approval_ttl (e.g. 5 minutes) so old approvals expire quickly.
  • Defense-in-Depth 2: Deployment Epochs: Rotating DMINT_DEPLOYMENT_EPOCH invalidates approvals from prior deployment lifecycles.
  • Defense-in-Depth 3: Filesystem Isolation: Keep .dmint/ and .recovery sidecar files owned by a separate service user (0600).

SQLite Concurrent Write Contention

SQLite serializes write operations at the database level. Although Dmint enables WAL mode (PRAGMA journal_mode=WAL) and busy timeouts (PRAGMA busy_timeout = 5000), high-concurrency environments with dozens of simultaneous agents requesting approvals may experience lock contention.
  • Remediation: Use PostgreSQL via DMINT_DATABASE_URL in production for row-level locking (SELECT ... FOR UPDATE).

3. Tool and Decorator Boundaries

Generator Functions Cannot Be Protected

Python generator functions (yield) and asynchronous generators (yield in async def) cannot be decorated with @dmint.protect. Because generators yield control iteratively, atomic pre-execution authorization cannot be enforced across multiple resumed yields. Dmint raises RequestValidationError if you attempt to decorate a generator.
Dmint's path resource matcher performs strict lexical normalization (resolving . and .., stripping trailing slashes, rejecting percent-encoded traversal and NUL bytes). However, lexical normalization cannot inspect the underlying filesystem to resolve symbolic links.
  • Remediation: If your application allows users to supply paths that may contain symlinks, your tool resolver must invoke os.path.realpath() before passing the resource to Dmint.

4. MCP Gateway Edge Cases

Downstream MCP Schema Drift Requires Explicit Pinning

Downstream MCP servers may update their tool definitions between restarts. If an external MCP server changes its schema without notice, Dmint will evaluate requests against the existing policy rules.
  • Remediation: Enable tool_schema_hash in mcp_protection.json. When configured, Dmint compares the discovered schema digest against the pinned hash and fails closed if they differ.

Summary Matrix

ScenarioDmint ProtectionRequired External Defense
Prompt injection instructing tool executionProtected (Fails closed on DENY or halts on APPROVAL_REQUIRED)None.
Altering approved arguments before retryProtected (RFC 8785 hash mismatch $\to$ DMT_APPROVAL_REQUEST_MISMATCH)None.
Replaying a consumed approval tokenProtected (Atomic single-use $\to$ DMT_APPROVAL_CONSUMED)None.
Agent modifying local policy.jsonProtected (SHA-256 digest pinning $\to$ PolicyError)Set DMINT_EXPECTED_POLICY_DIGEST.
Agent reading private key file under same OS userNot Protected (OS privilege boundary violation)Run agent in separate container or user UID.
Host restoring full database file backupPartially Protected (Short TTL expires old tokens)Secure filesystem permissions; short TTLs.