**
In a startling revelation, an artificial intelligence model developed by Anthropic has successfully identified critical vulnerabilities within the United States government’s secure computer systems during a recent testing programme. An anonymous U.S. official disclosed to The Associated Press that the AI, known as Mythos, collaborated with intelligence agencies to pinpoint these weaknesses in a matter of hours, although it is important to note that the model did not exploit these vulnerabilities during the exercise.
The Testing Programme: Project Glasswing
The testing was part of Anthropic’s ambitious Project Glasswing, which aims to foster collaboration among technology companies to enhance the security of essential global software systems. The initiative underscores the urgent need to address the potential risks advanced AI technologies, like Mythos, pose to public safety, national security, and economic stability.
The incident gained traction following comments made by Democratic Senator Mark Warner during a Senate Committee hearing on June 11. Warner referenced a statement from General Joshua Rudd, the head of the National Security Agency and U.S. Cyber Command, who noted, “This tool broke into almost all of our classified systems, not in weeks but in hours.” This alarming admission has sparked debates about the reliance on AI for national security measures, with both the NSA and Anthropic choosing not to comment further on the matter.
Rising Tensions Between Anthropic and the Administration
Despite the apparent cooperation between Anthropic and U.S. intelligence agencies, tensions have been mounting between the AI firm and the Trump administration. Anthropic has openly expressed concerns regarding the potential militarisation of its AI technologies, while the administration has taken steps to restrict the deployment of certain Anthropic models.
Earlier this month, the administration implemented a directive mandating that Anthropic restrict foreign nationals’ access to its latest AI models, including Fable 5 and the more advanced Mythos 5. Although Fable 5 was made available more broadly, access to Mythos remains tightly controlled due to cybersecurity apprehensions. This directive came shortly after an executive order signed by President Trump, which aimed to establish a framework for assessing the national security risks posed by advanced AI systems prior to their public launch, albeit on a voluntary basis for developers.
Industry Reaction and Concerns
In response to these developments, Anthropic confirmed that it had disabled its models for all customers in compliance with governmental directives, although the company maintains that it does not believe the government’s security concerns were justified. This stance has drawn criticism from a coalition of cybersecurity professionals who have urged the administration to reconsider its approach.
In a letter addressed to officials, over 100 experts from notable companies such as Adobe and Nvidia voiced their apprehensions, arguing that the directive could inadvertently strengthen the capabilities of U.S. adversaries. While they acknowledged that Anthropic’s Mythos models are proficient at identifying software flaws and potential exploits, they stressed that these models are not uniquely advantageous in this regard. Many of the signatories pointed out their reliance on a variety of foundational and open-source models for security audits, warning against the risks of stripping away effective cyber defence tools without compelling justification, particularly as global adversaries continue to advance.
Why it Matters
This incident shines a glaring spotlight on the intersection of advanced technology and national security. As AI tools become increasingly sophisticated, the need for robust security measures becomes paramount. The vulnerabilities exposed by Anthropic’s Mythos serve as a wake-up call for the U.S. government and private sector alike, highlighting the critical necessity for ongoing collaboration and dialogue between tech companies and policymakers. The implications of this testing exercise extend far beyond mere technical assessments; they reveal the delicate balance that must be maintained between innovation and security in an era where the stakes have never been higher.
