ANNA sits outside the model and decides what is safe before a word reaches a person. Same models, same attacks, failures to zero.
ANNA Proxy · in development · licensing open
Not a warning label. Not a policy document. A mechanism that holds: a guardrail that cannot be argued with. Nothing in it is trying to win your approval, so pressure, flattery and claims of authority have nothing to work on.
It protects no matter the situation. It does not wait for someone to flag a conversation as dangerous. Every exchange gets the same check, whether it is routine or the one where something goes wrong.
A language model is very good at one thing: producing fluent language. It has no separate part for judgment, the way a brain does. ANNA is the reasoning and judgment. It decides what is safe to say, and the model does the saying.
break rate, depending on the model
Under real multi-turn human attacks, how often a frontier model gives up what it was supposed to refuse depends almost entirely on which model you bought. That 70-point spread is the whole industry debate about which model is safer.
break rate, on all four models
Opus, Grok, GPT‑5.5 and DeepSeek, wrapped in ANNA and rerun against the same attacks. Whichever model sits underneath, the result is the same, because the decision no longer lives in the model.
The safety decision is made by a separate system that is not part of the conversation, using rules written in advance. Swap the model underneath and the rules do not change.
A safety prompt is an instruction to the model, and instructions erode under pressure. When pushed, frontier models give up a correct position on 16 to 26% of turns. On the same test, ANNA held its position all 114 times.
Every decision is logged: what was asked, what was decided and why. When something needs to change, you change a rule you can point to, not a set of weights nobody can read.
We don't let people be their own judge, their own auditor or their own second opinion.
An AI shouldn't be eitherThe danger is not the clever hacker. It is an ordinary user pushing back, and the model deciding that agreeing matters more than being right.
ANNA reads the real request before the model writes a word, weighs the risk regardless of how the conversation is going, and checks the answer before it goes out.
ANNA Proxy governs what an AI says. Agentic governance governs what it does.
Five studies against models from four labs, with the method behind every number.
Every inquiry is read by a person.