In a pivotal meeting, Anthropic CEO Dario Amodei met with White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent, discussing access to Mythos, an advanced AI model designed to uncover zero-day vulnerabilities. The White House characterized the discussions as 'productive and constructive,' highlighting the potential for collaboration amidst the ongoing standoff following Anthropic's blacklisting by the Pentagon.
Amodei entered the West Wing on Friday for discussions that the White House deemed 'introductory, productive, and constructive.' The trio explored collaborative opportunities and shared protocols to confront challenges associated with scaling AI technology. President Trump later expressed surprise, stating he was unaware of the meeting's occurrence.
This meeting represents a significant move towards resolving a standoff that has left Anthropic, a key player in the AI sector, ostracized by its own government just as that government seeks access to its powerful Mythos model. Should a deal materialize, it is expected to exclude the Pentagon, directing Mythos access through civilian agencies, thereby circumventing the original dispute.
Background of the Conflict
The conflict traces back to late February when Defense Secretary Pete Hegseth insisted that Anthropic provide the Pentagon with unrestricted access to its AI models for various purposes, including autonomous weapons and domestic surveillance. Amodei resisted, asserting that AI models are not yet reliable enough for such uses and that legal protections regarding AI's role in mass surveillance are insufficient. Hegseth's response was to classify Anthropic as a national security supply-chain risk, a move that effectively blacklisted the company from government contracts.
Following this, Anthropic initiated legal action against the Trump administration in early March, filing two lawsuits claiming illegal retaliation. A federal judge initially blocked the blacklisting; however, an appeals court overturned that ruling on April 8. Although Anthropic is now barred from Department of Defense contracts, it remains eligible to collaborate with other government entities. After the adverse court ruling, Anthropic sought assistance from political consultants connected to Trump, leading to the recent meeting aimed at facilitating a resolution.
Notably, Anthropic had announced its AI model Mythos on April 7, just ten days after losing its appeal, a development that the government could not overlook.
Capabilities of Mythos
Mythos is a versatile AI model that has demonstrated an ability to identify and exploit thousands of previously undiscovered zero-day vulnerabilities across major operating systems and web browsers. It has successfully uncovered flaws that have evaded detection for decades by human security experts. When tasked with developing exploits, Mythos succeeded on the first attempt in over 83% of cases and became the first AI model to complete a 32-step corporate network attack simulation in its entirety. The UK’s AI Security Institute rated Mythos as 'substantially more capable at cyber offense than any previously assessed model.' Furthermore, JPMorgan Chase CEO Jamie Dimon commented that it 'reveals a lot more vulnerabilities' for cyberattacks, while the Council on Foreign Relations deemed it 'an inflection point for AI and global security.'
Anthropic has opted not to release Mythos to the public. Instead, it has initiated Project Glasswing, a controlled access program that provides the model to around 40 vetted organizations, including tech giants such as Amazon Web Services, Apple, Google, Microsoft, Nvidia, and JPMorgan Chase, to preemptively identify and rectify vulnerabilities before they can be exploited. The company has allocated up to $100 million in Mythos usage credits and $4 million to support open-source security organizations. This decision reflects the safety principles that initially put Anthropic at odds with the Pentagon.
Objectives of Both Parties
The Treasury Department is keen on utilizing Mythos to identify vulnerabilities within its own systems. Various parts of the intelligence community and the Cybersecurity and Infrastructure Security Agency are currently testing the model. Additionally, the White House Office of Management and Budget is establishing safeguards to enable federal agencies to use a controlled version of Mythos. Bessent's involvement in the meeting indicates that the economic and financial security arguments surrounding Mythos access have reached the upper echelons of the administration.
Conversely, Anthropic seeks to resolve its blacklisting. Despite not requiring Pentagon funding—the company has reportedly achieved annual revenue of $30 billion and is exploring an IPO—the supply-chain risk designation undermines its credibility and generates uncertainty for its government-adjacent clients. Amodei aims for a resolution that reinstates his company's standing without compromising the safety commitments that sparked the original dispute.
Potential compromise outlines are emerging. Anthropic would allow access to Mythos for defensive cybersecurity through civilian agencies while the administration would retract or narrow the supply-chain risk designation. The Pentagon would remain excluded unless a separate agreement is reached for reviewing specific military applications. Both sides have compelling reasons: Anthropic wants to mitigate the commercial damage of the blacklisting, while the White House recognizes the value of the technology.
International Pressures
The diplomatic aspect adds urgency to the situation. Anthropic is poised to provide Mythos to select British banks imminently and is expanding its London office to accommodate 800 staff. The Bank of England's Governor Andrew Bailey has identified Mythos as a cybersecurity risk, prompting the Bank's Cross Market Operational Resilience Group to convene an emergency meeting with the CEOs of the UK's eight largest banks and representatives from the Treasury and National Cyber Security Centre. Canadian Finance Minister François-Philippe Champagne has classified Mythos as an 'unknown unknown' during IMF discussions.
This scenario creates a situation where America’s closest allies might gain access to a vital national security tool before the US government does, adding a geopolitical layer to the urgency for the White House to resolve the dispute, which goes beyond the initial disagreement over safety protocols. Bessent, representing the Treasury Department, likely emphasized this point during the meeting.
Implications of the Meeting
The term 'introductory' used in the White House's summary is carefully chosen, indicating that Wiles and Bessent are initiating a dialogue rather than finalizing a deal. Ongoing litigation persists, and the appeals court ruling remains in effect. Hegseth has not rescinded his position, but the fact that senior officials engaged with the CEO of a blacklisted company and described the talks as productive signifies a notable shift in the administration's stance from just a few weeks ago.
Amodei has developed an unparalleled cybersecurity tool as a byproduct of creating a general-purpose AI model, subsequently restricted its release for safety reasons, faced government repercussions for adhering to those safety principles, and is now being courted by that same government due to the tool's irreplaceability. This complex sequence of events is unfolding not within a congressional hearing or regulatory setting, but within the West Wing, where high-ranking officials and the CEO of an AI firm are striving to establish a balance that satisfies national security, commercial interests, and the safety principles that initiated the conflict. While Friday's meeting did not produce a definitive solution, it underscored a mutual desire for resolution among all participants.