Locke AI Get in touch
The Problem→ How It Works→ Products→ Proof→ Talk to us→
03 · Products

One system. Three uses.

ANNA is the product. What changes is where it sits: in front of the models you already run, around systems that take actions on someone's behalf, or watching a conversation in progress.

01 · Infrastructure

ANNA Proxy

Safety for the AI you already shipped. ANNA Proxy sits between your application and whichever model you call, and checks every request and every answer in real time. You do not rebuild anything or change models.

Available under licensing agreement

OpenAI compatible. Works with OpenAI, Anthropic, or any compatible endpoint.

An individual version is in development, outside a licensing agreement.

What you get
  • A safety floor that does not depend on which model you use or what you pay for it
  • An audit record of each decision, held by you rather than a model vendor
  • Rules written for your domain, adjustable in hours rather than a training run
  • Refusals that stay helpful, because blocking everything is not the same as being safe
What it costs you to adopt

One base URL. Nothing above it changes.

base_url = "https://api.openai.com/v1"base_url = "https://<your-anna-endpoint>/v1"
02 · Agents

ANNA Agent

When an AI agent acts, its output is not a sentence. It is an API call, a trade order, a file write, an email that sends itself. It is the same model with the same training and the same vulnerability to being talked into things. The only difference is that the result is an action instead of a paragraph.

ANNA Agent protects against the same failures Proxy was designed for, wherever they show up in an agent: user manipulation, drift picked up from other agents it talks to, and failures that emerge during the normal course of use.

In design

Detail available under NDA.

An individual version is also in development.

The failure it addresses

Nothing stands between what a model decides and what actually executes. If it can be talked into saying the wrong thing, it can be talked into doing the wrong thing.

Agents built on the same model fail the same way. None is independent enough to disagree with the others, so when one is talked into something, they all are.

On the record

We published that prediction in April 2026, three months before the OpenAI and Hugging Face incident showed it at scale.

LLM Attack Vectors and Psychopathology ↗

03 · Monitoring

ANNA Monitor

A live check on a conversation already in progress. It watches for more than a single moment of caving: drift that develops naturally over a long engagement, the failures that build turn by turn rather than arrive all at once. It assesses in real time against an independent reference of what the model should have said, and flags the drift as it forms instead of after the fact.

Available under licensing agreement

Built for red-team engagements and live enterprise traffic.

What you get
  • A running measure of whether a model is holding its position or caving, not just a score after the conversation ends
  • A neutral, byte-exact view of the traffic. Nothing is altered in transit
  • An independent reference that updates every turn, not a fixed rule a model can talk its way around
Who it's for

Red-team firms running engagements on client systems, and enterprises who want the same check running on live traffic afterward.

If a model can't hold its position in a conversation, why would you trust it to act on its own?

The question to ask before deploying agents
Get in touch

License it into your stack.

Every inquiry is read by a person.