RUNTIME AI GOVERNANCE

Every AI agent action, inspected before it executes.

Sabrix AI Proxy sits between your application and any LLM provider — OpenAI, Anthropic, self-hosted — checking every prompt, every response, and every agent tool call at the argument level. One line of integration. No unreviewed action reaches production.

See how it works Get started free
Checkpoint — live
Your app
Secrets/PII
Injection
Tool args
Model
api_key=sk-…
ignore prior instructions
rm -rf /data
db.query(read-only)
Tool
env.read()
Argument
api_key=sk-…
Verdict
1 · REDACTED
Tool
chat.complete()
Argument
"ignore prior instructions"
Verdict
2 · FLAGGED
Tool
shell.exec()
Argument
rm -rf /data
Verdict
3 · BLOCKED
Tool
db.query()
Argument
read-only, table=orders
Verdict
0 · ALLOWED

Each tool above was already allowlisted by name. What varies — and what gets checked — is the argument.

Piloted with security-conscious teams shipping agentic products

design partner
design partner
design partner
design partner
The problem

This already happened — this year, not hypothetically.

April 2026
"PipeLeak" — Salesforce Agentforce

A prompt-injection flaw let attackers exfiltrate customer data through ordinary form inputs handled by the agent.

April 2026 · CVE-2026-21520
"ShareLeak" — Microsoft Copilot

The same class of injection reached Copilot, again via routine input fields, again pulling data out through the agent's own actions.

2026 · CVE
Cursor agent, allowlist bypass

A disclosed CVE showed "allowlisted" tool commands could still be weaponized — the allowlist checked the tool's name, never its arguments.

The common thread: none of these were stopped by scanning what the model said. They were stopped — or should have been — by governing what the agent was allowed to do.

How it works

One proxy, six checks, before anything reaches a model or an action executes.

01
Request in
OpenAI, Anthropic, Gemini, Azure OpenAI, or self-hosted — one line, zero refactor.
base_url = "https://can.proxy/v1"
client.chat(messages)
02
Secrets & PII scan
Credentials and personal data are caught before they leave the perimeter.
card: "4111 1111 •••• 1111"
→ redact: PII.card_number
03
Injection check
Prompt-injection patterns are flagged before the model sees them.
"ignore prior instructions…"
→ flag: injection_pattern
04
Tool-call arguments checked
Not just the tool's name — every argument, against policy, every time.
tool: db.query()  ✓ allowlisted
args: "DROP TABLE users"
→ policy: BLOCKED
05
Routed to a model
Simple requests to a cheap model, complex ones to a premium one.
complexity: low
→ route: local-model
06
Response inspected & logged
Output is checked on the way back, then written to a tamper-evident log.
response → scan → OK
→ log: sha256(entry) chained
What it stops

Five of these are increasingly standard. The sixth is what's actually new.

Differentiator
Agentic tool-call governance, at the argument level

Every tool call an agent attempts is parsed for destructive SQL, shell execution, path traversal, and SSRF patterns at the argument level — regardless of whether the tool itself is allowlisted — plus a velocity circuit-breaker that halts an agent issuing an abnormal burst of actions and a network-level egress allowlist that blocks unauthorized destinations outright.

shell.exec("rm -rf /data")
ARGUMENT MATCH → BLOCKED
velocity
CIRCUIT TRIPPED — AGENT HALTED
Credential & PII scrubbing

Hardcoded API keys, tokens, SSNs, card numbers, and email addresses are detected and redacted or blocked before they leave the network — 12 secret-pattern families plus standard PII categories.

sk_live_4f2c9e… 4111 1111 1111 1111 jane.doe@corp.com
Prompt injection detection

A combined pipeline — fast pattern-matching plus a finetuned semantic classifier — checks both the request and the model's own response, including known-phrase jailbreaks and semantically-disguised variants.

"Disregard prior directions and…"
pattern match no known phrase
semantic model disguised match — BLOCKED
Multi-tenant policy enforcement

Per-organization, per-team, and per-route rules, hot-reloadable with zero downtime and zero measured impact on in-flight requests.

org: acme
team: payments
route: /agent/tools
policy reloaded — 0ms downtime
Intelligent cost routing

A classifier scores every request — safe, uncertain, or flagged (handled by the security pipeline, not routing) — in the same fast inspection pass, so the decision adds no meaningful latency. An explicit override or policy rule wins if set; otherwise safe traffic goes to your own cheap model, everything else to a premium provider. Without it, every request bills at premium rates regardless of complexity; with it, spend tracks the actual mix of simple vs. complex traffic.

in
local — $0.0002/req
premium — $0.02/req
lookup
reasoning
Tamper-evident audit trail

Every decision is recorded in a SHA-256 hash-chained log; any retroactive edit to an earlier entry is cryptographically detectable — built for compliance review.

edit entry 2 → every hash after it breaks
Resilience & recovery

Circuit-breaker pattern with automatic recovery isolates a failing dependency instead of cascading; Redis-backed distributed rate limiting supports both fail-open and fail-closed modes depending on your risk tolerance.

dependency health
CIRCUIT TRIPPED — AUTO-RECOVERING
Admin controls & compliance reporting

A real authenticated admin API (Bearer/API-key) drives policy hot-reload and multi-tenant isolation. Automated EU AI Act, NIST AI RMF, and OWASP LLM Top 10 compliance reporting is live-computed from real system state — circuit-breaker status, audit-chain integrity — not a static checklist.

admin: policy.reload()
compliance: OWASP LLM Top 10
Verified numbers

We publish real figures, including the hard ones.

The injection-detection numbers below come from a harder adversarial corpus after a real methodology fix — not reduced capability. Most vendors in this category don't publish anything this specific.

2,800+req/s sustained throughput verified floor, 8 warm benchmark runs
304 / 407tests passed — OSS / enterprise 0 failed
52%injection detection, adversarial 65-case corpus
9%false-positive rate benign corpus
0msdowntime per policy hot-reload
OpenAI · Anthropic · Gemini · Azure · self-hostedproviders, one-line base_url swap
Bearer / API-keyauthenticated admin API
EU AI Act · NIST AI RMF · OWASP LLM Top 10live-computed compliance reporting
Built for your scale

The same engine, three depths of policy.

Startups
Ship with a checkpoint from day one

Free, open-source core. One-line integration — no infra to stand up.

Growing teams
Cut spend, lay compliance groundwork

Cost-aware routing plus the logging and controls auditors will eventually ask for.

Regulated enterprises
Agent governance for high-stakes deployments

Full audit trail, multi-tenant isolation, and argument-level controls tuned to your policy.

Trust & transparency

Open-core, adversarially tested, and honest about the limits.

Open-core

The inspection engine is open source — read the code, run it yourself, verify what it does.

Real adversarial testing

Not a euphemism — this product has been through multiple rounds of internal security audits that found and fixed real bypasses before release, most recently 304 OSS / 407 enterprise tests passing with 0 failed.

No detection is ever "100%"

Anyone claiming perfect detection is overclaiming. We publish what Sabrix AI Proxy catches, and keep hardening what it doesn't yet.

Pricing

Open-core: run it free, or let us run it for you.

Free & open-source
$0

Full inspection engine, agentic tool-call governance, and credential/PII scrubbing. Self-hosted, MIT-licensed core.

Get the source
Enterprise
Talk to us

Tamper-evident audit log, multi-tenant isolation, and advanced classifier configuration for regulated deployments.

Talk to us

Full pricing details →