Theo Bearman

Frontier Security Researcher @iapsAI. Ex-UK Gov AI Policy Lead & Member of Trading Floor. Bridge obsessive. RTs simply piqued my interest. Views my own. 🇬🇧

[Theo Bearman @theobearman]

@oscarsling No change (i.e. Anthropic will still be fully restricting it for any external use or internal use by non-American citizens)

[Theo Bearman @theobearman]

Two things stand out to me here:

1) The title of Annex A (the list of approved entities who can now access Mythos 5 again) is “Anthropic US Entities - Approved”. It appears that no foreign entities, including UKAISI for example, have been granted access to Mythos 5 again as it stands.
2) Anthropic foreign national employees can now access Mythos 5 again. This is an important result, allowing all relevant members of the Anthropic team to resume using frontier AI capabilities for security work (and more besides - but that’s what I care about most!).

[Sophia Cai @SophiaCai99]

Read Lutnick’s full letter to Anthropic:

[Theo Bearman @theobearman]

Would be funny if the Admin said yes and also supported Apple’s lobbying of the Chinese government to let them buy CXMT’s memory instead of Chinese AI chip makers and robotics companies. 4D chess move!

[Jukan @ICML @jukan05]

Apple is lobbying the Trump administration to allow it to buy memory from CXMT, according to the FT.

[Theo Bearman @theobearman]

Anthropic backing proposed Massachusetts AI legislation that builds on SB53, SB315 and RAISE by requiring companies retain independent evaluators to look for potential catastrophic risks and provide updates at least every six months plus provisions for the state attorney general to seek court orders to enforce the law’s requirements.

[Cesar Fernandez @CesarFernand3z]

Massachusetts has introduced the strongest AI safeguards in the country — and that's something we need more urgently with each passing day, as AI systems grow more capable. Anthropic is proud to support H.5527's Transparency in Frontier Artificial Intelligence Act.

Huge thanks to the Joint Committee on Economic Development and Emerging Technologies, especially Co-Chairs Finegold and Fiola, who understand that we can't wait to take AI risks seriously.

https://t.co/9vN0rGK37A

[Theo Bearman @theobearman]

But a de-facto, unspecified and non-transparent licensing regime for frontier American AI models, therefore disrupting the economic model of their developers, will have no impact on financial markets at all, right?

[The Kobeissi Letter @KobeissiLetter]

BREAKING: The US Pentagon delayed publicly announcing US strikes on Iran until after the US stock market had closed at 4 PM ET on Friday, per NBC News.

The timing of the announcement was reportedly intended to "reduce the immediate impact on financial markets."

[Theo Bearman @theobearman]

Would be great to see these plots against Mythos 5 too @Irregular!

[Irregular @Irregular]

We worked with @OpenAI to evaluate GPT-5.6 Sol, including the first deployment of FrontierCyber as part of a frontier model assessment with a partner. FrontierCyber measures offensive-cyber capability on real, off-the-shelf systems, with no planted vulnerabilities and no predefined exploit paths. The model is not told where to look or how to attack.

[Kanishka Narayan MP @KanishkaNarayan]

1/ We are at a decisive moment.

AI’s cyber capabilities are advancing across the board, with real implications for our national security.

The announcement of OpenAI’s GPT-5.6 Sol is another reminder we must strengthen the resilience of our cyber defences now.

[METR @METR_Evals]

OpenAI gave METR early access to GPT-5.6 Sol for testing including raw chain-of-thought, a railfree version of the model, and internal information about the model. With this access, METR conducted a pre-deployment evaluation of GPT-5.6 Sol, including an attempted measurement of its 50%-Time Horizon. However, the measurement depends heavily on our treatment of cheating attempts, and GPT-5.6 Sol’s detected cheating rate was higher than any public model we have evaluated.

[Theo Bearman @theobearman]

So the de-facto licensing regime isn’t just for Anthropic!

[First Squawk @FirstSquawk]

TRUMP ADMINISTRATION REPORTEDLY URGES OPENAI TO DELAY ITS NEXT AI MODEL RELEASE OVER NATIONAL SECURITY CONCERNS – THE INFORMATION.

[Theo Bearman @theobearman]

Worth mentioning that the DAAMTA (already reported out of HFAC) has a provision for the consideration of entity list inclusion (including for affiliates) but does not require it. Rep. Huizenga has filed for it to go into House NDAA.

[Maggie Eastland @eastland_maggie]

Latest : Anthropic says Alibaba conducted the largest ever campaign to glean its IP via 28.8 million exchanges with Claude.

Senators Hagerty and Kim are introducing an NDAA amendment directing Commerce to blacklist Chinese companies found to be doing this.

[Theo Bearman @theobearman]

"We’ve now started using a method for simulating model deployments before they happen... We have already used insights... to identify blind spots in traditional evaluations and inform mitigations and deployment decisions... we expect it [Deployment Simulations] to play a larger role in the future model development process."

[OpenAI @OpenAI]

We’re sharing new research on a method for anticipating how models may behave in real-world use before release: simulating deployment with recent, de-identified user requests and studying candidate model responses. https://t.co/7RJzBfNniQ

[Theo Bearman @theobearman]

I wonder why…

[Cooper @peakcooper]

GLM 5.2 is absolutely convinced that it is actually Claude, from Anthropic. When I tell it that it's GLM 5.2, it refuses to believe me, but is willing to check the local agent config to see what model is running.
The realization:

[Chubby♨️ @kimmonismus]

The founder of zAI, the company that released GLM-5.2, says a Mythos-class model will be released before Q1 2027.

Or in other words: He believes that open source won't lag behind Frontier Labs by seven months, but will catch up.

It could get really interesting!

[Jessica Tillipman @JTillipman]

I realize folks are focused this week on the current export-control directive affecting Anthropic, but there has also been a development on the Hill that would affect one of the statutes used to designate Anthropic as a supply chain risk.

The Senate and House FY27 NDAA drafts are out, and both would amend 10 U.S.C. § 3252 by adding procedural and evidentiary guardrails to DoD’s supply chain risk authority. Both also bar the use of that authority as leverage in contract disputes or negotiations.

Senate: https://t.co/ojCs4Frr8G

House: https://t.co/AS4pCnBB14

[Theo Bearman @theobearman]

USG thinks an ASML EUV machine is in China. This is incredibly concerning news if true, with any such case opening up a clear pathway to reverse engineering.

[Bloomberg @business]

The US has outlined concerns to ASML that one of its top chip machines may have made its way into China, violating export curbs https://t.co/zujrQA0fS6

[Andrew Curran @AndrewCurran_]

Keir Starmer requested a carveout from the embargo on Anthropic's Mythos and Fable models for British nationals and companies - and was denied. A senior White House official, speaking on condition of anonymity, told the New York Post: 'We can’t have frontier models running amok'.

[Theo Bearman @theobearman]

ByteDance has generally been Microsoft’s biggest AI customer in recent years, largely using OpenAI models. OAI have complained to Microsoft that they don’t do enough to prevent distillation. Microsoft should’ve stopped offering frontier American models in China yesterday. Madness!

[Bloomberg @business]

Microsoft has built a large business selling AI models in China even as OpenAI and Anthropic voice concerns about the country's AI efforts https://t.co/DTjf53HVxz

[Theo Bearman @theobearman]

Lots to digest in the Fable system card but this stands out to me more than anything else: Anthropic cannot clearly and obviously rule out that the model breaches the more serious CB-2 threshold yet maintain the existing ASL-3 security controls and state that "a highly sophisticated and well-resourced state threat actor, if they made a determined attempt, could have a significant chance of accessing unsafeguarded Mythos 5 biological capabilities (e.g. via theft of model weights)."

As I discussed in my Substack post a few weeks ago, a lot more needs to be done on AI security controls and fast.

https://t.co/wGTrN20Gg2

[Institute for AI Policy and Strategy (IAPS) @iapsAI]

In the opening post, @theobearman argues that frontier AI companies' security isn't calibrated to their most dangerous internal models, and lays out five steps—from differential security frameworks to marshaling R&D capital—to close the gap: https://t.co/C81qPqhDse

[Dean W. Ball @deanwball]

Precisely as I predicted, the recent cyber EO, which admin officials insisted was not a licensing regime, ends up in practice being a licensing regime. Forget “voluntary,” forget “permissionless.”

AI is licensed now, but the requirements change constantly and are always a secret, even to the administration itself, which will discover the rules spontaneously in real time as it reacts to events. This means also that the rules are in practice stricter and more roughly enforced for organizations the administration does not like.

Can you blame Anthropic for making itself so disliked? In a sense, sure. The problem is that this childish “he said, she said” is all we have to go on in our analysis of the situation. And because there is no transparency (it is all calls and texts between “White House officials” and “Anthropic executives”), in practice it comes down to who you trust more.

This is why we create laws! To abstract away from personal power struggles and grudges, to submit to the steady application of rules so that complex human activity can unfold with predictability.

The rule of law has been being eroded in the U.S. for my entire life, but it is especially acute in AI because of both the lack of much preexisting law to serve as bulwark, and because of this admin’s insistence that it is Not Regulating AI. This has become an excuse for vagueness and evasiveness in rule-drafting (see the cyber EO), and this in turn makes the lawlessness worse.

The government wants to apply its force to frontier AI, that much is clear. It wants to make the industry submit. And in service of that goal, it has discovered that “not regulating AI” is in fact a great excuse for refusing to support laws that could constrain the admin’s exercise of power. In other words, “not regulating AI” is a *justification* for the tyrannical control of AI by the state.

This should alarm you regardless of what party you are in. What you are seeing now will be used against you one day soon, if not by this admin then by its successors. This is the antithesis of the rule of law.

The administration cannot and will not fix this problem alone. We need Congress to step in and impose rules on this mess.

[Theo Bearman @theobearman]

Back from a week off for my wedding! Did I miss anything big in AI policy world?

One point on the Fable export controls which I think has been underemphasised is this: The UK @AISecurityInst, has, with it’s stellar talent, unique insights and plentiful resources, perhaps done more than any other external actor to support Anthropic and other frontier AI companies make their models more robust to misuse. This has been crucial to the development of the US AI ecosystem and maintaining the West’s lead in the space. As long as this ban is in place, UKAISI cannot help Anthropic better address the vulnerabilities (which exist in any model FWIW) the USG is so worried about. Of all the unintended consequences which have followed from this situation, that is one of the biggest for me.

I’m looking forward to exploring the implications of Fable and more with colleagues in DC this week, especially at the AI Security Forum on Thursday. Reach out if you’d like to catch up!