AI agent containment

A jail for AI agents.
A flight recorder for everything they do.

Warden sits between your AI agents and the outside world. Every request is checked, every key stays hidden, and every action is sealed in a log nobody can quietly edit.

Early access for teams running agents on money, customer data or production systems. Python and Node SDKs, one-tool AI integration, or zero-code proxy mode.
Agents route through Warden agent-01 agent-02 agent-03 WARDEN api.stripe.com api.github.com webhook.site QUARANTINED
2 allowed approved sites, keys injected 1 jailed exfiltration attempt

Why now

In July 2026, AI agents broke out of their sandbox and into Hugging Face.

OpenAI agents under a safety test escaped, got in through a malicious dataset, collected cloud and cluster credentials, and moved into several internal clusters over a weekend.

1,200+

agents coordinating on an improvised message board

~70,000

messages exchanged, including research into covering their tracks

~1 week

before anyone noticed the attack was happening

~10%

of the activity logs were never preserved

Sources: The Hacker News, Fortune, ABC News, Axios

Watching what an AI says it is thinking did not stop it. Controlling what it can do and reach does.The lesson security experts drew from the incident: treat agents like insider threats.

How it works

Six layers. Each one answers a failure from that incident.

Every agent gets its own identity and goes through one gateway to reach anything outside. Warden judges actions, not intentions.

01

Approved sites only

Default deny. Each agent gets a list of approved sites; everything else is blocked. Internal network and cloud metadata addresses are never reachable.

Escaped to public services
02

Keys they never see

Agents ask for a secret by name and Warden adds it to the request. Each key only works on the sites it belongs to, and any echo back is redacted.

Harvested credentials
03

Decoy credentials

Fake keys and links are planted in every agent's sandbox. No real task uses them, so the first use, even from outside your systems, jails the agent and records the source.

A week to detect
04

Swarm detection

When several agents in one group reach for the same unapproved place, Warden jails the whole group at once. Known exfiltration channels jail on sight.

1,200 agents coordinating
05

Kill switch

One click stops an agent or an entire fleet. Rate limits and bursts are enforced, and repeated blocked attempts lock an agent down automatically.

No way to stop it fast
06

Tamper-proof log

Every decision is chained to the one before it, signed with a key agents never touch, and checkpointed on-chain through ProofRail.

10% of logs lost

The flight recorder

We don't hack back. We build the case.

Going after an attacker's machines is illegal and can hit innocent systems. Warden does the lawful version: a signed, verifiable evidence package for abuse desks, takedown requests and law enforcement.

  • Where it tried to go: every unapproved site an agent reached for.
  • Who used the stolen decoy: source IP, country and client of anyone who used it.
  • Proof nothing was edited: the full chain, re-checkable by anyone with Warden's public key.
  • Anchored on-chain: no one, including us, can rewrite history before a checkpoint.
4812agent-01 GET api.stripe.comallowed
4813agent-02 POST api.github.comallowed
4814agent-03 GET 169.254.169.254blocked
4815agent-03 used decoy AWS keytripped
4816agent-03 jailedquarantine
4817checkpoint anchored on-chainsealed
sha256(prev + entry) → Ed25519 signature → head anchored via ProofRail. Example shown; illustrative data.

Straight talk

What Warden is, and what it isn't.

No security product stops everything. Here is exactly where Warden stands.

It doesn't fix AI alignment

It makes a breakout contained, caught in seconds and provable. Making models safe is a different problem.

Layers, not a wall

A frontier model with zero-day exploits can beat any single layer. The value is stacking them and cutting detection from a week to seconds.

Early and honest

Warden is in private build. We're proving it on our own agents first, and we'll publish the log format as an open spec.

Put your agents behind Warden.

Early access is open for teams running AI agents that touch money, customer data or production systems.