Defense in depth

No single layer is ever enough.

That is the oldest rule in security, and the reason a serious system is built in layers: a breach only happens when the holes line up across every one of them (the Swiss-cheese model). So we do not ask you to trust any single control. Here are our layers, mapped to the functions every security program is measured by, the NIST Cybersecurity Framework and the CIA triad, and each one states its own limit. Reading the limits is the point.

The layers (NIST CSF)

Identify

Discovery + BorderBuilt

Inventory the MCP servers you actually have configured, and flag config-level risk the server scan cannot see: unpinned loaders (the rug-pull surface), hardcoded secrets, internal endpoints.

The honest limit

Sees what you submit; it does not crawl your whole environment for agents you never told it about.

Protect

ExecutiveStrong

A deterministic allow / require-approval / deny gate with no model in the blocking path, so a jailbroken agent cannot argue past it. Per-agent policy and budgets, and an inline gateway so an agent is governed by pointing its config at us, not by wrapping its code.

The honest limit

Cooperative: it governs calls that route through us. An adversary with code execution on the same host can skip an in-process SDK, which is why the gateway and the record exist as further layers.

Detect

Intelligence + PoliceBuilt

Deterministic detectors for tool poisoning, malicious arguments, artifacts, and output injection, plus a standing patrol that scores each actor for wide, exfiltration-shaped coverage between its calls.

The honest limit

Pattern-based detection is bypassable by design; the patrol is advisory and reads only instrumented actors. A novel attack can pass a signature.

Respond

CorrectionsBuilt

A human-pulled kill switch and an automatic circuit breaker, revocation, and a signed incident bundle. Findings export to your SOC as OCSF, so we are a slice in the stack you already run.

The honest limit

The kill switch stops governed calls, not an adversary who left the governed path. Actor-scope action is never autonomous, on purpose.

Recover

JudiciaryBuilt

Contest a decision and have a reversal recorded as visibly as the accusation, on the same tamper-evident ledger, and re-pin the version. The offline retro-hunt reconstructs a campaign after the fact.

The honest limit

Recovery is of the record and the governed subject; recovering a system already breached out of band is not ours to do.

Govern

Legislature + ConstitutionStrong

A published, versioned, non-retroactive corpus you can read before you are judged, and an institution bound by a record no one, including us, can rewrite.

The honest limit

Laws change by our release today; an independent legislative process is still to come.

The CIA triad

ConfidentialityStrong

Enforcement runs in your own process; no tool description, output, or session state leaves your infrastructure. Baselines are per-customer, never pooled. A deny redacts its own reasoning from the agent.

IntegrityStrong

Ed25519 receipts, a hash-chained ledger, RFC 6962/9162 Merkle inclusion and consistency proofs, and a daily external timestamp. Anyone verifies it offline with our published package. This is the leg most tools cannot match.

AvailabilityBuilt

The kill switch fails safe, not open, when its state cannot be read; an unreadable stop routes to a human rather than allowing blind. Rate limiting and a customer-chosen fail mode default to tightening, never opening, under an outage.

Layers raise the cost of a breach; they never reach 100 percent, which is the whole premise. The field itself says a sound AI-agent architecture needs at least one deterministic enforcement layer that does not rely on model reasoning, and that is our core. Defense in depth is honest by construction: it assumes every layer has holes and builds the next one anyway. See the standing limits on the threat model and how the whole thing is organized on the institution page.