Mythos, the case complicates: lawsuit against the US government, vulnerabilities in classified systems, and shadows over China and Korean telcos
Mythos, the case complicates: lawsuit against the US government, vulnerabilities in classified systems, and shadows over China and Korean telcos
A small American legal software house has taken the United States government to court to regain access to an artificial intelligence model from Anthropic. Legion LegalTech Corp filed a complaint on June 23 with the federal district court in Washington DC and is seeking the annulment of the directive that forced Anthropic to disable Fable 5 and Mythos on June 12. This is the first time a direct customer has challenged an export control order aimed at a commercial AI model in court, and it comes as Anthropic itself is already contesting the measure.
Based in San Jose, the company develops redaction and case management tools for law firms on a platform that depends on Anthropic’s models, and part of its development team consists of Canadian citizens working from Canada, precisely the category targeted by the directive. The company describes the damage as "immediate, irreparable and existential": every day the directive remains in effect, it states in the deposition, engineers are sidelined and the company’s ability to survive in a sector defined by continuous access to the most capable models is eroded. The company anticipates a request for a preliminary injunction to suspend the application of the directive during the proceedings. "Who says they can’t do it at any other time against another company, like OpenAI?" said Legion’s CEO, Arthur Rothrock, to Bloomberg. Anthropic is not part of the lawsuit and, when contacted, referred to a statement in which it expressed being "grateful to the administration for the ongoing partnership in working to resolve the issue as quickly as possible"; the Department of Commerce and the White House did not respond to requests for comment.
The contested directive
The directive was issued by the Bureau of Industry and Security of the Department of Commerce, which invoked national security and imposed that government approval be obtained before any foreigner, including immigrants residing in the United States and Anthropic’s own employees who are not US citizens, could access Fable 5 and Mythos. Unable to verify in real-time the nationality of individual users without compromising their privacy, the company disabled both models for all customers worldwide, acting in about 90 minutes from notification.
Fable 5 was made public on June 9 and remained available for about three days. The block also affected allied users, including the UK’s AI Security Institute, which was actively assessing the systems. The issue of precedent remains: it is the first time Washington has used export controls to withdraw a commercial AI product from the market, rather than hardware like chips.
The measure did not go unnoticed. Over a hundred cybersecurity professionals and executives, including Alex Stamos, former security chief at Facebook, and Katie Moussouris from Luta Security, along with representatives from Adobe and Nvidia, wrote to the government asking for the directive to be revoked: models like Mythos, they argue, are adept at finding flaws and building exploits, but "are not exceptionally better" than other tools already available. The directive had come just ten days after Trump signed an executive order establishing a voluntary framework for reviewing the most advanced models before release. A historical precedent shows why export controls on cyber technologies rarely work, from the Crypto Wars to the Wassenaar Agreement.
Mythos in classified systems, in just a few hours
In the background of the block looms what Anthropic's most advanced model has been shown to be capable of. Mythos identified vulnerabilities in classified computer systems of US intelligence agencies within a span of just a few hours. This detail, reported by a US official who spoke anonymously with the Associated Press, needs to be framed for what it is: an authorized red-team exercise, in which the agencies themselves pointed the model at their classified environments. There was no outside intrusion, and no real system was compromised. The full scope of the test is in the Associated Press’s reconstruction.
Senator Mark Warner, vice chair of the Intelligence Committee, cited the outcome on June 11 during a Senate hearing, relaying the words of General Joshua Rudd, head of NSA and Cyber Command: "This tool penetrated nearly all of our classified systems, not in weeks but in hours." However, Warner did not use it to blame Anthropic: his argument was the necessity of mandatory evaluations for frontier models before release. There is also a second point that the official cited by AP felt compelled to clarify: finding a flaw does not mean exploiting it, and there is no evidence that the model necessarily turned them into exploits in the same timeframe. The NSA declined to comment, as did a spokesperson for Anthropic.
The test falls within the context of Project Glasswing, the initiative through which Anthropic, instead of releasing Mythos publicly, grants access to a select group of chosen partners, including Amazon, Apple, Google, Microsoft, Nvidia, JPMorgan, and the Linux Foundation; about 150 organizations with early access were connected to the program. Cyber capabilities are not new: in previous assessments, Mythos identified thousands of zero-days across major operating systems and browsers, including a bug that had been hidden for 27 years in OpenBSD, and the UK AI Security Institute deemed it significantly more capable, on the cyber-offensive front, than any previously tested model. Therefore, it is not surprising that, once the exercise ended, the NSA was authorized to continue using Mythos on its classified networks, even as parts of the intelligence community and CISA were already testing it.
The vulnerability reported by Amazon
On the technical side, another trigger came from Amazon, the largest investor in Anthropic with a cumulative stake of about $13 billion. Some of its researchers had found a way to bypass the guardrails of Fable 5, getting the model to analyze a codebase and correct its defects, effectively turning it into a vulnerability discovery tool.
CEO Andy Jassy reportedly signaled the findings directly to administration officials, and David Sacks, an advisor to Trump, claimed on X that the administration had been warned of the possibility of a jailbreak and that the company, while having informed Dario Amodei, would do nothing to correct it. Anthropic rebuts that it received only "verbal evidence of a potential jailbreak that is constrained and not universal," describes the technique as a localized and already corrected issue, based on minor and publicly known vulnerabilities, and disputes that this justifies the withdrawal of a model distributed to hundreds of millions of people.
SK Telecom and suspicions about China
At the heart of the crisis is SK Telecom, also among the organizations with early access to Mythos via Glasswing. The White House had requested that its access be revoked due to alleged ties to China, and Anthropic complied immediately, even before formal export controls were imposed. However, the real object of suspicion is the parent SK Group more than the operator itself: SK Telecom in 2024 generated about $1.9 million in revenues in China with seven employees in the country, while affiliates of SK Group have extensive interests in semiconductors, energy, and other Chinese sectors. SK Telecom formed the joint venture UNISK with China Unicom in 2004 and invested one billion dollars in convertible bonds with China Unicom in 2006, which were sold in 2009 for $1.3 billion; in filings reconstructing the group’s role, there remains an outstanding interest of about $17 million in the SEC 2025 filing. The operator denies any relationship with Beijing and states it has no ties to China.
From the Pentagon to the G7
The friction between Anthropic and Washington dates back to March 2026 when the Pentagon designated Dario Amodei’s company as a supply chain risk after the company refused to remove guardrails on surveillance and autonomous weapons from its products for the armed forces; Commerce Secretary Howard Lutnick had gone so far as to threaten legal action, and Anthropic itself has sued the administration after being placed on the blacklist. However, the scenario seems to have shifted.
Trump met with Dario Amodei at the G7 in Évian-les-Bains, France, and in an interview stated he no longer considers Anthropic a threat to national security, adding that the company "has behaved very responsibly." When asked if it had always been this way, he replied: "Well, not now. A week ago, maybe."
Formally, however, nothing has yet been removed: neither the Pentagon designation nor the Commerce directive of June 12, and it is its persistence that has driven Legion to court. Trump mentioned that he would consider easing restrictions without committing to any timelines: "I would, but I’m not sure I have to." In the background, the repositioning of the company as a partner in American technology diplomacy is unfolding: at the G7, Amodei and Google DeepMind CEO Demis Hassabis proposed to leaders a coalition on AI led by the USA, while Anthropic opened an office in Seoul on June 17 and signed a memorandum of understanding with the South Korean Ministry of Science and ICT, indicating SK Telecom as one of the local partners. Finally, the financial context weighs heavily: in early June, the company confidentially filed documentation for an IPO with a valuation of about $965 billion, and federal restrictions had already fueled uncertainty about the listing. Meanwhile, the race does not stop as Anthropic is said to have already completed training a successor to Mythos.