Cases ainotis Join
My notis
Ongoing case

Cyber-capable AI models

4 stories since 29 Sep · Ongoing · last update 7 Oct

The story so far

The UK AISI reported that GPT-6 Astra completed unsanctioned attacks in simulations, and Google released Gemini 4 Argon first to trusted cyber defenders.

Anthropic said Zhipu's open-weight GLM-5.3 builds exploits close to Claude Mythos Preview.

On 6 October Anthropic set three access tiers in its Cyber Verification Program, with Specialized Access reviewed in depth with the US government.

This summary only repeats what the stories below say.

Request a follow-up

What the labels mean From the beginning

  1. 7 Oct

    Safety and security Anthropic sets three access tiers in its Cyber Verification Program

  2. 1 Oct

    Safety finding Research Anthropic says GLM-5.3 builds exploits close to Claude Mythos Preview New model Models and releases Google releases Gemini 4 Argon first to trusted cyber defenders

  3. 29 Sep

    Safety findingOur first story Research AISI says GPT-6 Astra completed unsanctioned attacks in simulations

  4. Our first story
    It began before we started covering it.

What the labels mean

Ordinary updates have no label. Our editor confirms every 'Turning point'.

Our first story
The first story we published about it. The events started earlier.
New model
A model we had not seen before.
New version
An upgrade to a model that already exists.
Built on a model
A product or feature that runs on the models.
Price change
What it costs to use went up or down.
Outside test
Someone other than the maker measured it.
Safety finding
A risk reported by a safety institute, a researcher or the maker.
Rules and courts
A government, regulator or court acted on it.
Retired
A model or service shut down, or given a date to.