Mythos Meltdown: Anthropic's Troubles Escalate

The ongoing dispute with the White House poses significant and potentially severe consequences for the United States' artificial intelligence sector.

Two weeks have passed since Anthropic was compelled to withdraw its advanced Mythos-class AI models from service, following an abrupt ultimatum issued by the Trump administration on a Friday evening. In response, the company swiftly dispatched a team of executives to Washington, D.C., though subsequent updates have been notably scarce, leaving the situation unresolved.

Anthropic has repeatedly refused to comment on the status of these discussions throughout the week, citing a lack of new information. However, the silence itself is telling. Following 14 days of intense negotiations, uncertainty persists regarding the potential return of Anthropic’s most potent AI models, and whether President Trump might extend this directive to other companies utilizing comparable technologies. Each passing day without a resolution exacerbates the severity of the situation, impacting not only Anthropic but the broader U.S. AI industry.

The Trump administration's export control order, issued on June 12th, mandated that Anthropic restrict "any foreign national" from accessing Mythos 5 and Fable 5, citing national security concerns. This comprehensive ban applies to all non-U.S. citizens, whether located within or outside the United States, including those employed directly by Anthropic. To date, Anthropic has determined that maintaining these models offline remains its sole viable course of action.

The precise reasons for the continued deadlock between Anthropic and the administration remain unclear. A contributing factor may be the absence of an established framework for applying export controls specifically to AI systems. While most manufacturers of dual-use products — civilian technologies with potential military or defense applications — can typically assess compliance through a standardized checklist during production, Anthropic is navigating a complex bureaucratic process to apply these regulations from fundamental principles.

Typically, this type of export control review can span months, or even years, and is usually finalized before a product's market debut. However, as previously reported by The Verge, the U.S. Department of Commerce reportedly tested Fable 5 prior to its release without raising any objections. A source close to the negotiations indicated that Anthropic had deemed its models safe for public availability. The agency's intervention seemingly occurred only after an individual, reportedly Amazon CEO Andy Jassy, highlighted a method for ostensibly bypassing Fable 5’s safety guardrails, leading to the entire regulatory process being dramatically expedited into just a few days.

Katie Moussouris, founder and CEO of Luta Security, reviewed a report on the Fable 5 vulnerability at Anthropic's behest and considers the concerns significantly exaggerated. In a blog post, Moussouris elaborated on how researchers managed to bypass guardrails designed to prevent Fable 5 from identifying exploitable security flaws—a capability considered one of the most alarming aspects of the unrestrained Mythos 5. While the model would decline requests to "review code for security issues," it would respond to prompts to "fix this code" when followed by specific manual instructions, potentially causing it to reveal vulnerabilities it was programmed to withhold.

Moussouris contends that this functionality should not have provoked such a drastic governmental response, arguing instead that it represents an indispensable tool for AI-assisted coding. She stated, "Defenders need to be able to ask AI to fix the bugs in a file, explain why the fix matters, and write tests that confirm the patch works. That is not a guardrail bypass. It is the most valuable thing an AI model can do for defensive security: executing the find, fix, and test loop defenders run every day."

Over the past week, Anthropic co-founder Tom Brown has reportedly taken over from CEO Dario Amodei in negotiations with the Trump administration, joined by Sarah Heck, the company’s public policy chief, as reported by Wired. Despite this change, the discussions appear to be progressing slowly, with actual advancements remaining questionable.

Irrespective of the causes for this prolonged delay, Anthropic has suffered a considerable setback. Prior to these extended negotiations, the company was recognized as a rare AI entity demonstrating a clear trajectory toward profitability. Its Mythos-class models, with input tokens priced at twice that of its less powerful Opus 4.8, were anticipated to significantly enhance revenue in preparation for an impending IPO. Moreover, Mythos’ cybersecurity capabilities had seemingly begun to mend relations with the Trump administration, following months of legal and rhetorical contention.

The revenue generated by Mythos is crucial for Anthropic to cover its recently acquired computing resources, including a reported $15 billion annual agreement with SpaceX for data center access, and to maintain a positive public image ahead of its IPO. Two of Anthropic’s major shareholders, Google and Amazon, have diligently sought to remain in the Trump administration's favor, suggesting their probable dissatisfaction with the current situation.

Concurrently, the sluggish pace of negotiations has fostered a power vacuum within the global AI market. This stems not only from the Mythos shutdown but also from the U.S. government’s demonstrated readiness to restrict American AI systems perceived as risky—a concern relevant to models from several other U.S. companies, including OpenAI, Google, and Microsoft, which may present similar risks to Mythos. Consequently, there has been an emerging demand for non-American AI solutions from other nations. As cybersecurity expert and Corridor Chief Product Officer Alex Stamos remarked to The Verge last week, "One of America’s champions is being kneecapped by the US government while we’re in a race with the Chinese. It’s just incredibly stupid."

With each passing day, the predicament for these companies intensifies. Their AI models are nearing Mythos-level capabilities, potentially triggering further export control orders; indeed, OpenAI’s GPT-5.5 Cyber recently surpassed Mythos 5 in specific benchmarks. The Trump administration has also reportedly requested OpenAI to postpone the launch of GPT-5.6 due to security concerns, with intentions for the government to individually approve each customer. Both Anthropic and OpenAI have impending IPOs, while China continues to widen its lead in the global AI competition.

Ironically, this administrative order follows months during which the Trump administration advocated for dismantling AI safeguards and regulations, making it one of President Trump's first significant regulatory actions in this domain. However, numerous cybersecurity leaders have collectively asserted that if regulation is necessary, this approach is fundamentally flawed. Despite the Trump administration’s pledges to reverse Biden-era AI regulations, it appears to have, in many respects, reclaimed and even expanded regulatory authority in the AI landscape.