An official from the United States, speaking anonymously to The Associated Press, revealed on Tuesday that Anthropic’s artificial intelligence technology had uncovered security weaknesses in U.S. government computer systems. This discovery was made during a test exercise employing Anthropic’s Mythos model.
The official noted that the collaboration with U.S. intelligence agencies allowed Anthropic to identify these vulnerabilities swiftly. However, the model did not exploit these weaknesses in the short time it took to find them. The testing was a part of Anthropic’s Project Glasswing, which aims to protect crucial software from potential threats posed by advanced AI models like Mythos.
Anthropic’s Initiative and Government Collaboration
Project Glasswing, initiated by Anthropic, brought together various tech giants and organizations to address the risks associated with AI technology. The initiative’s goal is to safeguard public safety, national security, and economic stability from potential threats posed by advanced AI capabilities.
During a Senate Committee hearing on June 11, Democratic Senator Mark Warner of Virginia mentioned the tests. He reported that the Mythos model was able to break into many classified systems rapidly, a statement based on information from General Joshua Rudd, head of the NSA and U.S. Cyber Command.
Government and Corporate Responses
Following these revelations, neither the NSA nor Anthropic provided further comments. Nonetheless, the collaboration highlighted existing tensions between Anthropic and the Trump administration. Anthropic has voiced concerns over the military’s potential use of its AI, while the administration has placed restrictions on some of the company’s models.
Recently, the administration issued a directive limiting foreign nationals’ access to Anthropic’s advanced models, Fable 5 and Mythos 5. This directive aligns with an executive order signed by President Trump, which establishes a framework for assessing the national security risks of advanced AI systems before their release.
Industry Reactions and Future Implications
In response, Anthropic disabled its models for all customers to comply with the government’s directive, although it expressed doubts about the necessity of such measures. Meanwhile, a group of cybersecurity leaders has urged the administration to reconsider its stance, arguing that limiting access to advanced AI tools could benefit U.S. adversaries more than it would help them.
Over 100 cybersecurity experts, including leaders from companies like Adobe and Nvidia, signed a letter asserting that while Anthropic’s Mythos models are adept at identifying software flaws, they are not uniquely so. They cautioned against removing top-tier cyber defense tools without substantial justification, especially as global adversaries continue to enhance their capabilities.
As the debate continues, many in the cybersecurity community emphasize the importance of maintaining robust defensive capabilities in the face of evolving technological threats.
