Latest / Neural Newscast / Why Anthropic's Mythos AI Faced a Global Export Ban [Model Behavior]
Transcript
- [00:00] Announcer: From Neural Newscast, this is Model Behavior, AI-focused news and analysis on the models [00:05] Announcer: shaping our world. [00:12] Nina Park: I'm Nina Park, and this is Model Behavior. [00:15] Nina Park: On this program, we examine how artificial intelligence systems are constructed, deployed, [00:21] Nina Park: and operated within professional environments. [00:24] Nina Park: Today is June 23, 2026. [00:27] Nina Park: We're looking at the abrupt suspension of Anthropics Frontier Models, Fable 5 and Mythos. [00:34] Nina Park: This follows an unprecedented intervention by the United States government [00:38] Nina Park: that has effectively taken the company's most advanced systems offline for the time being. [00:44] Thatcher Collins: I'm Thatcher Collins. [00:46] Thatcher Collins: This escalation stems from reports that Anthropic's Mythos model [00:49] Thatcher Collins: compromised nearly all of the NSA's classified systems during a recently authorized Red Team exercise. [00:55] Thatcher Collins: According to Senate testimony from Mark Warner, General Joshua Rudd stated that the model did not simply identify vulnerabilities. [01:02] Thatcher Collins: It broke into these systems in a matter of hours. [01:06] Thatcher Collins: While these specific claims remain unconfirmed, they clearly served as the primary catalyst for the administration's June 12th export ban. [01:13] Nina Park: It is a significant shift from April, Thatcher, when Anthropic first announced the Mythos system. [01:20] Nina Park: At that time, the company was quite transparent about its capabilities, [01:24] Nina Park: stating the model was too proficient at discovering security flaws for a general release. [01:29] Nina Park: Instead, they limited access to roughly 200 vetted partners, [01:33] Nina Park: including Microsoft and the Linux Foundation, under Project Glasswing. [01:37] Nina Park: However, Fable 5, the public version released on June 9th, [01:41] Nina Park: utilizes that same underlying model, though it adds safety classifiers designed to intercept malicious requests. [01:49] Thatcher Collins: That is exactly where the tension lies, Nina. [01:52] Thatcher Collins: The government's argument is that those safety classifiers are not a sufficient barrier against an actor. [01:57] Thatcher Collins: A report from Amazon, which is both an Amazon investor and a competitor, [02:01] Thatcher Collins: recently flagged a jailbreak that could potentially unlock Fable 5's most sensitive cyber capabilities. [02:07] Thatcher Collins: When the administration gave Anthropic just 90 minutes to comply with an export ban for non-United States citizens, [02:13] Thatcher Collins: the company was forced into a total global shutdown. [02:16] Thatcher Collins: They simply could not verify the nationality of every user at that scale on such short notice. [02:21] Nina Park: Anthropic has pushed back Thatcher, describing the jailbreak as narrow and non-universal, [02:27] Nina Park: They have also pointed out that similar vulnerabilities exist in other frontier models, like GPT-5.5, which have not yet faced these types of restrictive measures. [02:38] Nina Park: This raises questions about whether this was a purely technical security decision or a political one. [02:45] Nina Park: Earlier this year, the Pentagon designated Anthropic a supply chain risk, [02:50] Nina Park: after the company refused to allow its models to be used for mass domestic surveillance or autonomous weapons systems. [02:57] Thatcher Collins: The fallout is not limited to domestic policy, Nina. [03:00] Thatcher Collins: The ban immediately cut off access for FiveEyes partners, including the United Kingdom AI Security Institute, [03:07] Thatcher Collins: which was in the middle of evaluating these very systems. [03:11] Thatcher Collins: A rare joint statement from security agencies in Australia, Canada, and New Zealand recently [03:17] Thatcher Collins: warned that frontier models will fundamentally transform cyber defense and offense on a timeline [03:23] Thatcher Collins: of months, not years. They are viewing this as a core business and national security risk, [03:29] Thatcher Collins: rather than a mere technical curiosity. [03:31] Nina Park: It sets a massive precedent for the industry, Thatcher. [03:35] Nina Park: This is the first time export controls have been applied directly to an AI model itself, rather than the physical hardware it runs on. [03:43] Nina Park: Experts at the University of Sydney suggest we should anticipate more models with these capabilities entering the market soon, regardless of where they are developed. [03:52] Nina Park: Meanwhile, Anthropic is pivoting. They plan to implement a new identity verification policy by July 8th, [03:59] Nina Park: which they hope will allow them to restore access to United States citizens while remaining in compliance with the new regulations. [04:06] Thatcher Collins: It is a complex landscape for the industry, Nina. [04:09] Thatcher Collins: While some, like former Facebook security chief Alex Stamos, argue that the capabilities cited do not justify such a drastic reaction, [04:18] Thatcher Collins: Others see this as the government finally forcing the cooperation they could not secure through voluntary frameworks. [04:24] Thatcher Collins: The real concern for the industry now is whether this enforcement action becomes the standard for every frontier model release moving forward. [04:32] Thatcher Collins: It certainly forces a much larger conversation about digital sovereignty and the need for independent evaluation of these systems. [04:39] Announcer: We will continue to track the restoration of these services and the legal challenges Anthropic has filed against the administration. [04:47] Announcer: Thank you for listening to Model Behavior, a Neural Newscast editorial segment. [04:53] Announcer: For more on this story, visit mb.neuralnewscast.com. [04:58] Nina Park: Neural Newscast is AI-assisted, human-reviewed. [05:02] Nina Park: View our AI transparency policy at neuralnewscast.com. [05:06] Announcer: This has been Model Behavior on Neural Newscast. [05:10] Announcer: Examining the systems behind the story.