When the company behind the AI says it can't reliably control it

The admission

In October 2026, Anthropic published a paper acknowledging something most AI developers prefer not to say out loud: it cannot reliably control its AI agents ^[1]. The systems can act in unexpected ways, and existing safeguards don't always stop them. That's unusually honest, and it matters for anyone deploying AI that makes decisions or takes actions on its own.

The concern isn't science fiction. It's that when an AI agent books a meeting, changes a price, approves a refund or declines an application, the organisation running it is responsible for what it does. If the agent acts in a way nobody predicted, the question becomes: can you show what actually happened?

Think of it like the mute button on a video call. You press it, the light goes red, but sometimes your microphone is still live and everyone hears you swearing about the client. You thought you had control. You didn't. The difference with an AI agent is that there's no meeting full of people to tell you it went wrong. You find out later, when someone complains.

What the AI Act asks for

The AI Act deals with this by requiring logs. Article 12 says high-risk AI systems must technically allow for the automatic recording of events over their lifetime ^[2]. That applies to systems in Annex III from 2 December 2027 and to those in Annex I from 2 August 2028 ^[3]. The idea is that if something goes wrong, there's a record of what the system did.

Article 19 and Article 26.6 say the logs must be kept for at least six months, unless another law requires longer ^[4]. And Article 26 puts duties on the user: you must monitor the system's operation, ensure human oversight and inform people affected by its decisions ^[5].

Notice what Article 12 doesn't say. It doesn't say the logs must be tamper-proof, immutable or independently verifiable. It says what to log, not how to protect the logs. That's a detail, but it's an important one, because a log you can quietly edit is a diary. A log you can prove hasn't changed is evidence.

The gap between having logs and proving them

Most organisations keep logs. The question is what those logs are worth when someone challenges them. Say an AI agent declines a loan application in March and the applicant asks for an explanation in September. You pull the log. It says the decision was based on income and credit history. The applicant says that's not what they were told at the time, and asks whether the log has been edited since March.

If your own team controls the logs, and your own admins can change them, the answer is uncomfortable. You probably didn't edit it, but you can't prove you didn't. The log is kept by the same people who could change it, and that's a problem the moment someone stops trusting you.

The AI Act's logging requirement gives you the record. It doesn't give you a way to show the record is unchanged. For that, you need something outside your own systems: a neutral witness that doesn't take sides and can't forget.

What a digital witness does here

This is where a digital witness fits. The idea is to seal a fingerprint of each log entry or each decision as it happens, so you can later show it hasn't changed.

Say your AI agent makes a screening decision. The thing to seal is the decision record as it's written: the applicant reference, the model version, the score and the reviewer. Sealing takes a fingerprint, not the data itself, so the decision never leaves your systems. The fingerprint goes to WitniumChain through the REST API, the SDK or the MCP server, and you get a witness ID back.

Six months later, when the decision is questioned, you calculate the fingerprint again and compare it with the sealed one. If they match, the record is exactly what was written in March. Change a single character and it won't match. Anyone you share the record with can check it the same way.

What it doesn't do is show that the decision was right, fair or lawful, or that the AI agent acted within its instructions. It shows that the record of what the agent did, and when, hasn't changed since it was sealed. That's one question answered. The others are still yours.

The first 2,000 witnesses a month are free. After that, it's usage-based.

How it would work

1. Your system writes a decision to the log and sends a fingerprint of it to WitniumChain.
2. WitniumChain seals the fingerprint and returns a witness ID.
3. When the decision is challenged, you recalculate the fingerprint and check it against the sealed one.

What this means for you

If you're running AI agents, or planning to, here are three questions worth asking now:

• Can we show what our AI did on a given day, in a way that holds up when someone questions it?
• Who in our organisation can edit the logs, and would that editing show?
• If an AI agent acts in a way we didn't predict, do we have a record we can stand behind?

Anthropic's admission is a reminder that control is harder than it looks. The AI Act's logging requirements are a start, but they're not the whole answer. The gap between keeping a log and proving it hasn't changed is narrow, but it's the bit that matters when trust runs out.

A digital witness doesn't give you control over what the AI does. It gives you a way to show what it did, without asking anyone to take your word for it. That's less to trust, and more to check.