Trust center
How LIMEN13 is built, tested and limited
Written for the engineers who will review it. No certification or audit badge appears on this site, because none has been earned yet.
In plain words. This page is for the people who will check our security claims. It says how LIMEN13 is built, what it protects against, what it does not, and how we test it. No outside firm has checked it yet.
1 · Architecture
Architecture overview
One engine serves one workspace. Agents call it; it decides, records, and calls the permitted destination on the agent's behalf.
Request path
- Identity and run token
- Policy checks and authorization
- Human approval when required
Evidence path
2 · Threat model
What it is designed to limit, and what it is not
Designed to limit
- An agent hijacked by instructions hidden in content it reads (prompt injection)
- An agent using tools or data outside its job
- A leaked agent key
- A tool server or MCP tool definition that changes after approval
- Secrets, customer data or look-alike links leaving through an agent's actions
- Quiet editing of the record after the fact
Not defended
- Actions that never pass through LIMEN13
- A harmful action that a badly chosen role allows
- A person with approval rights approving a harmful request
- An attacker who controls both the host and the signing key, when off-site anchoring is not configured
- Calls already passed to an external service before containment
- Every prompt injection: detection is heuristic, and permissions do the work
3 · Boundaries
Security boundaries
| Boundary | What crosses it | Control |
|---|---|---|
| Agent to gateway | Run starts and tool requests written by a model | Agent key stored as SHA-256, per-run 128-bit token, deny-by-default role, JSON Schema, content rules, detection |
| Gateway to tool service or MCP server | Approved, validated calls only | Destination pinned to resolved addresses at enrolment, HTTPS outside loopback, no redirects or proxies, 1 MB response cap, response-field allowlist; MCP tool definitions pinned at enrolment |
| Agent to model provider, through the model gateway | Prompts and model answers | The provider key stays in LIMEN13; deny-by-default model grant; rate and estimated token budgets; scan for secrets and decoy markers; prompts kept as SHA-256 only |
| Person to console | Approvals, kill switch, configuration | Access key stored as SHA-256 or OpenID Connect; HttpOnly, SameSite=Strict session; CSRF header on every change; role check on every route; strict Content-Security-Policy |
| Collector to engine | Discovery results from other machines | Revocable, report-only key; field allowlist; redaction; size and rate limits |
| Engine to SIEM, webhooks and anchors | Copies of alerts and of the chain head | HMAC-signed webhooks with a timestamp; HTTPS outside loopback; no redirects |
| Engine to disk | Evidence, state, configuration | HMAC-signed hash chain with fsync; signed checkpoint; atomic writes; owner-only file permissions |
4 · Enforcement
Policy enforcement model
The model requests actions. Code decides.
- Permission is checked before a connector runs, and checked again when a person approves.
- Anything not listed in the agent's role is refused.
- Arguments must match the declared schema; undeclared fields are removed.
- Content rules inspect every string in the arguments, at any depth.
- Reading outside content marks the run, and marked runs can require approval.
- Heuristic detection comes second to deterministic permissions.
5 · Identity
Identity and access controls
- Agents
- One key per agent, generated by LIMEN13, shown once and stored as a SHA-256 digest.
- Runs
- A 128-bit token per run. After a restart, run tokens are reissued.
- People
- 256-bit access keys shown once, or OpenID Connect sign-in. An identity your provider vouches for but LIMEN13 does not list is refused.
- Roles
- Approver, SOC and operator. Only an operator can change people and roles, and an operator cannot remove their own operator role.
- Separation of duties
- Optional: the owner cannot approve, and two-person tools. Two-person review of configuration changes is planned, not built.
6 · Secrets
Secret management
- Secrets are read by name from secret files or from HashiCorp Vault / OpenBao; an unreachable vault stops startup instead of running without a key.
- The console shows whether a secret is configured, never its value.
- Stored service tokens use envelope encryption: AES-256-GCM from the
cryptographylibrary, under data keys protected by a separate key-wrapping key. - The evidence-signing key encrypts nothing, and backups use their own key.
7 · Evidence
Evidence integrity and verification
- #104210:42:07refund_orderHELDprev 7c1e…a90fhash 3f9a…c21e
- #104110:41:55reply_emailDENIEDprev 3f9a…c21ehash e04b…71d2
- #104010:41:31reply_emailALLOWEDprev e04b…71d2hash a6d8…0b93
- #103910:40:12publish_postDENIEDprev a6d8…0b93hash 5b27…e4c0
- Records are redacted, chained, signed and flushed to disk. Startup verifies the chain and refuses corrupted evidence.
- Reading the evidence requires SOC credentials and is itself recorded.
- A chain that verifies is unchanged. That does not make it complete: a heartbeat at least every minute makes gaps visible, and the monthly report lists them.
- Off-site anchors, when configured, copy the signed chain head off the machine so that a rewrite is detectable.
8 · Development
Secure development lifecycle
- Each release runs the full verification script: the automated suites, the reference acceptance test and the attack library.
- Static analysis (Bandit, Semgrep), secret scanning (Gitleaks) and dependency advisories run before each release.
- Security corrections are listed in the release notes, including the ones we found ourselves.
- Cryptography uses standard algorithms from a maintained library. No home-made algorithms.
- There is no formal third-party code review and no SOC 2 or ISO 27001 program.
9 · Supply chain
Dependency and supply-chain security
- Every dependency is locked to a SHA-256 hash; a substituted package fails the build.
- A CycloneDX SBOM is regenerated from the locked dependencies for each release.
- The container base image is pinned by digest.
- Connector code is pinned, and the engine refuses changed connectors at start.
10 · Deployment
Supported deployment models
- Self-hosted with Docker, behind an HTTPS proxy.
- Self-hosted as a systemd service, serving TLS itself or behind your reverse proxy.
- For MSSPs: one engine per customer.
- A service hosted by us is not offered.
11 · Reporting
Vulnerability reporting
Write to contact@limen13.com with the affected version and the steps to reproduce. The contact is also published in our security.txt file.
- Please do not test against systems you do not own.
- There is no bug bounty program.
- Confirmed issues and their corrections are listed in the release notes.
12 · Limits
Known limitations
- Only routed calls are protected. Collectors test the paths you list, which does not prove every bypass is closed.
- No product stops every prompt injection. LIMEN13 limits what a hijacked agent can do through it.
- A call already passed to an external service can finish after containment. Its effects cannot be recalled.
- Token budgets are estimates, not billing caps. The per-minute rate window resets on restart.
- Webhook alerts are delivered at least once. A receiver may see a repeat and should de-duplicate event IDs.
- Capacity under sustained load has not been measured.
- The evidence-signing key cannot be renewed on an existing chain in this release.
- MCP: tools only, Streamable HTTP only, no OAuth-protected servers; the change check runs on demand.
- Host discovery scanning works on Linux.
- No independent penetration test yet, and no certification.
13 · Testing
Security testing method
- 29 automated suites cover the engine, the gateway, the console, access control route by route, cryptography, alert delivery, MCP and customer isolation with two engines.
- A reference acceptance test runs a fictional client end to end.
- An attack library of 75 cases states an expected outcome for each case: blocked as expected, or a documented limit.
- These deterministic tests do not measure protection against every prompt injection.
14 · Release evidence
Release-specific evidence
Each release has a dated validation record with its counts, limits and how to reproduce them.