On today's episode of Decoder, the guest is Hayden Field, senior AI reporter for The Verge. Often when Hayden joins the show, it is because something has gone wrong in the world of AI. Last weekend, that something involved a striking combination of Anthropic, the Trump administration, and the company's new AI model, Fable 5.

On Friday, less than a week after Anthropic made Fable available to the public, the US government announced export controls on the new model along with the underlying Mythos model that powers it. The restrictions barred foreign nationals, including those employed by Anthropic inside the United States, from using these systems. Anthropic responded by taking both Fable and Mythos offline for all users, citing concerns that it could not reliably limit access while meeting the order's requirements.

The episode explores this complicated situation in detail. Hayden recently published a thorough account on The Verge detailing the sequence of events from Friday and the intense discussions over the weekend as both parties worked to clarify what occurred and identify possible resolutions. The goal of the conversation is to trace the timeline and examine its broader implications.

Verge subscribers receive exclusive ad-free access to Decoder on their preferred podcast platforms. Those who are not subscribers can sign up through the provided options.

The matter remains unresolved. As of Tuesday during the recording, Fable stayed offline, with Claude displaying a notice above the chat window that Fable 5 is currently unavailable. Hayden explains that regardless of whether Fable returns this week, the effects of the government's conflict with Anthropic extend well beyond the immediate issue, influencing the technology sector and the emerging US approach to AI oversight.

A notable irony surfaces in the discussion: Anthropic has long warned that AI could soon reach levels of capability that pose serious risks and has advocated for timely government regulation. Now that such involvement has arrived, the company finds the execution less than ideal.

Observers, including the Chinese government, are closely monitoring whether the United States develops a substantive safety framework for AI or simply deploys regulatory tools to pressure companies that resist administration preferences. The episode describes the entire episode as highly unsettled.

The transcript has been lightly edited for length and clarity.

Hayden Field, senior AI reporter at The Verge, returns to Decoder.

Thanks. It is good to be here.

Chaos tends to follow your appearances, Hayden. That pattern continues.

It does. The pattern holds that events unfold on Fridays, and this instance followed the same rule.

Regulatory actions on AI often produce frantic Friday developments that spill into weekend analysis, a pattern that has repeated during the current administration with Monday efforts to restore order.

Exactly. The pace is intense, with discussions continuing through the weekend and into Monday without immediate resolution. The drama remains high.

The podcast airs on Thursday after a Tuesday recording, leaving room for further developments. The larger questions about US AI regulation and its impact on American companies globally stay open and will recur with future model releases.

The discussion begins by clarifying the models involved. What are Claude Mythos and Fable, and how are they connected?

Mythos serves as the base model behind both Mythos 5 and Fable 5. Fable represents a new release that starts at version 5 without prior public iterations. It functions as a safeguarded variant of Mythos 5. Earlier previews emphasized Mythos as a potential cyber tool too risky for broad release until additional protections were ready, limiting initial access to enterprises and government entities focused on defense.

Last week Anthropic launched Mythos 5 as the first full version and simultaneously introduced Fable 5 as the first public offering in that class. Both share the same core but Fable includes extensive added safeguards intended to prevent misuse of capabilities that Anthropic itself described as highly concerning.

Fable applies security measures to Mythos. Its launch sparked debate because researchers found the protections sometimes blocked legitimate testing by routing advanced queries back to older models.

The strictness quickly became a topic of online commentary and humor.

The contrast is striking: criticism arose over overly tight controls, yet government action treated the model as sufficiently hazardous to warrant removal. What triggered the shift?

Independent testers initially reported that the safeguards performed well against various attempts. Midweek, Amazon researchers shared findings about a possible jailbreak with Anthropic. Discussions between the teams assessed its severity. Reports indicate Amazon CEO Andy Jassy raised the issue with Treasury Secretary Scott Bessent, prompting administration officials to contact Anthropic with a demand to address the matter within 90 minutes.

Anthropic sought clarification on the specific concerns during follow-up calls. The administration expressed frustration over response timing, and once the deadline passed without full agreement, export controls were imposed. These barred access by any non-US citizens, whether employees, enterprise users, or individuals. Anthropic concluded that immediate compliance required taking the models offline entirely.

Subsequent statements from Anthropic indicated ongoing cooperation efforts, including weekend virtual meetings and in-person discussions in Washington. CEO Dario Amodei participated remotely. Industry observers noted widespread concern, with one source remarking that Beijing was likely amused by the developments.

Cybersecurity figures issued a public statement emphasizing that even supporters of regulation viewed the current approach as flawed. The episode underscores that effective oversight requires methodical and consistent processes, qualities absent from the recent sequence of events.

Questions arise about how non-experts in senior roles navigated the technical details, leading to rapid escalation without deeper evaluation of the reported issue.

The lack of specialized knowledge contributed to a binary response. While any potential vulnerability merits attention, unfamiliarity with AI systems can produce overly rigid reactions without full context on how such models operate.