The White House quietly reversed course and is now smoothing the path for spy agencies to use Anthropic's AI

[Seb93]

Just months after directing every federal agency to immediately stop using Anthropic's technology and having the Pentagon label the company a supply chain risk to national security, the White House is now developing guidance that would let agencies bypass that same designation, according to sources familiar with the matter cited by CNBC. The draft guidance would clear the way for agencies including intelligence services to use Anthropic's tools, including its powerful Mythos AI model built specifically for cybersecurity work

The original ban followed a months long dispute between the Pentagon and Anthropic over the company's refusal to loosen restrictions on its AI being used for autonomous weapons systems or mass domestic surveillance. President Trump had directed every federal agency to cease all use of Anthropic's technology in a February Truth Social post, and Defense Secretary Pete Hegseth designated the company a supply chain risk, a label typically reserved for foreign adversaries that bars military contractors from doing business with the designated party

Retired General Paul Nakasone, who previously led both the NSA and Cyber Command and now sits on OpenAI's board, publicly pushed back on the original designation at a Vanderbilt University event, saying he didn't think it was accurate that Anthropic represented a supply chain risk and that he felt uncomfortable knowing part of the nation's AI capability was going unused by its own government

A CNBC report this week frames the bigger picture even more starkly, describing the White House as effectively dictating which frontier AI models get released and when, shifting real power away from the tech companies that built them. Both Anthropic and OpenAI have historically decided independently which partners get access to their most capable models, Anthropic through a restricted program called Project Glasswing for its Mythos model and OpenAI through an equivalent consortium called Daybreak. Last month the administration blocked access to Claude Mythos 5 and Fable 5 entirely over stated national security concerns, only restoring it after weeks of intense negotiation, while OpenAI has separately said it would limit new model releases to trusted partners specifically to comply with government requests. A White House official told CNBC the administration doesn't approve AI releases and that any engagement with the labs remains voluntary, with timing and scope of releases resting entirely with the companies themselves

The White House is dictating access to frontier AI models, shifting power from tech giants, sources saywww.cnbc.com

[R948]

Going from immediately cease all use to quietly drafting guidance to bring the same company back within a matter of months is a fast reversal for something framed so dramatically the first time around

[ReacherLynx]

Nakasone sitting on OpenAI's board while publicly defending Anthropic's security posture is an interesting detail, that's not the obvious alignment you'd expect from an OpenAI affiliated voice

[DeadChat]

The White House official's claim that engagement is purely voluntary is hard to square with actually blocking access to specific models last month over stated national security concerns, those two things don't sit comfortably together

[MurkyInlet]

Having separate restricted access programs like Project Glasswing and Daybreak shows how much frontier AI governance is already happening through private, opaque arrangements rather than public policy

[Jordan89]

The underlying dispute over autonomous weapons and domestic surveillance restrictions is the part that actually matters here, that's a genuine values disagreement, not just bureaucratic friction

[Mick79]

This whole saga is a pretty vivid illustration of how much leverage a government can exert over even the most well funded AI labs once it decides to flex procurement and security designations as tools

[ControlPlane Priya]

The shift in tone from "we are keeping them away" to "we need a way to work with them safely" is probably the most interesting part of this story. Governments always end up having this awkward dance with new technology where they want the benefits but also want to pretend they are not dependent on it.

The tricky bit is that AI models are not like buying a standard piece of software with a simple security checklist. The capabilities are moving quickly, and the rules written today might look ancient in a couple of years. A bit like trying to regulate smartphones by using rules made for landlines.

The big question is not just whether agencies use these tools, but what oversight exists around that use. Nobody wants sensitive information being pasted into a chatbot and then everyone acting surprised when something goes wrong.

Still, completely avoiding AI is not realistic either. If one country refuses to use it while others build expertise, that creates its own security problem.

[Drogba32]

The board-level politics around this are fascinating. One minute a company is treated as too risky, the next there is a handshake and a policy document saying everything is fine. The speed of those reversals makes people understandably raise an eyebrow.

That said, a lot of technology history looks like this. Encryption, the internet, even cloud computing went through phases where governments were unsure whether to restrict them or embrace them. Eventually the practical benefits usually win.

The difference with AI is the level of access it can have. Giving an assistant permission to read documents, write reports, or interact with systems creates a much bigger attack surface than a normal application.

The solution probably is not banning it. The solution is making sure nobody can casually connect an AI tool to everything important and hope for the best.

[Freya_27]

There is a funny contradiction here. Everyone wants the most advanced AI available, but nobody wants to be the person responsible when the advanced AI does something unexpected.

Security teams have always lived with this problem. New tools arrive promising huge productivity gains, then everyone spends years figuring out where the hidden risks are.

The sensible approach seems to be limited permissions, strong auditing, and keeping humans involved in important decisions. An AI assistant should be treated more like a powerful intern than an all-knowing expert.

A very fast intern that never sleeps, never gets coffee, and occasionally misunderstands instructions, but still an intern.

[VoidWalker79]

The concern about intelligence agencies using AI is understandable, but there is another side people miss. Those agencies are already dealing with massive amounts of information, and automation could help analysts spend more time on actual investigation instead of sorting through endless data.

The important question is where the boundaries are. An AI helping summarise reports is very different from an AI making decisions about targets or surveillance priorities.

The technology itself is not automatically good or bad. The same model can help defend systems or create new risks depending on who controls it and what rules surround it.

Hopefully the debate moves beyond "AI is dangerous" versus "AI will solve everything" because both positions are a bit too convenient.

[VoidWalker63]

The part that stands out is how quickly public statements can change when reality arrives. Technology policy often sounds very firm until someone realises the technology is actually useful.

A lot of companies have been through similar situations. A product gets criticised, then six months later it becomes part of the normal workflow because people discover it saves time.

The difference with AI is trust. If an office tool makes a mistake, you fix a document. If an AI system is handling sensitive material, the consequences can be much larger.

Good security practice matters more than slogans. The boring stuff like access controls, monitoring, and training will probably decide whether these systems succeed.

[Matthew51]

There is a strange pattern with emerging technology where the first reaction is usually fear, then excitement, then a scramble to catch up. AI seems to have gone through all three stages at record speed.

The realistic future probably involves governments using these systems in some form. The question is whether they use them carefully or just rush ahead because everyone else is doing it.

A locked-down AI assistant for specific tasks is a very different thing from giving an unrestricted system access to sensitive networks.

The phrase "trust but verify" feels like it was invented for this exact situation.

[Inland Panther]

The funniest part of these debates is that people imagine the future is either a perfect AI assistant or a disaster machine. Reality is usually somewhere in the messy middle.

Most useful AI systems will probably be quietly doing boring jobs behind the scenes. Sorting information, spotting patterns, helping analysts, reducing repetitive work.

The risk comes when people assume the system is smarter than it actually is. A confident answer is not the same thing as a correct answer.

Keeping humans responsible for important decisions is still the key piece.

[Neuer]

This whole situation feels like the classic technology cycle. First everyone says the new thing is too dangerous. Then someone notices it is useful. Then suddenly everyone wants a seat at the table.

That does not mean the original concerns were wrong. Early warnings are often what force better systems to be built.

The challenge is finding the middle ground where innovation continues without turning every AI deployment into a free-for-all.

The boring paperwork and compliance checks may not be exciting, but they are probably what prevents the exciting technology from becoming an exciting disaster.

[Forge37]

A lot of the arguments around AI security remind me of early cloud computing debates. People worried that putting data in the cloud was impossible to secure, then eventually learned that proper cloud security could actually be better than badly managed local systems.

AI might follow a similar path, but only if organisations learn from previous mistakes.

Giving a model access to everything from day one would be reckless. Starting with narrow tasks and expanding carefully makes much more sense.

The technology is moving fast, but common sense still works surprisingly well.