Anthropic’s Mythos AI Unveils Security Weaknesses in US Systems, Officials Confirm
Anthropic s Mythos AI found flaws – During a recent security assessment, Anthropic’s Mythos AI model swiftly uncovered weaknesses in classified US government systems, according to a senior official. The discovery, which occurred within hours of the test, has raised alarms among officials about the potential risks posed by advanced artificial intelligence. The official, who requested anonymity, revealed that Anthropic had partnered with US intelligence agencies to evaluate the system’s capabilities through a joint initiative. This collaboration, part of a broader effort to assess AI’s impact on national security, highlighted how quickly the model could identify vulnerabilities in highly sensitive infrastructure.
Project Glasswing: A Collaborative Security Initiative
The testing was conducted as part of Anthropic’s Project Glasswing, a program designed to assess critical software for weaknesses that could be exploited by AI systems. This initiative brought together technology firms and government agencies to preemptively address threats from models like Mythos, which are capable of posing significant risks to public safety, economic stability, and national security. According to the official, the goal was not only to detect flaws but also to understand how quickly such systems could compromise data integrity in real-world scenarios.
Senator Mark Warner, a Democrat from Virginia, referenced the findings during a June 11 hearing before the Senate Committee on Banking, Housing, and Urban Affairs. He emphasized the urgency of the situation, stating that the AI had demonstrated an alarming ability to breach classified systems. “This tool infiltrated nearly every classified system in our government, accomplishing what others might take weeks to achieve,” Warner said, citing General Joshua Rudd, the head of the National Security Agency (NSA) and US Cyber Command, as the source of the information.
A New Era of AI-Driven Cyber Threats
The test results underscore the growing sophistication of AI in cyber warfare. While the model identified vulnerabilities, it did not immediately exploit them, the official clarified. This distinction is crucial: the AI’s ability to detect flaws signals a potential threat, but the actual execution of attacks would depend on further development and access to specific systems. Nonetheless, the speed at which Mythos uncovered weaknesses has prompted officials to reconsider the security protocols surrounding AI deployment.
Anthropic’s collaboration with US agencies has been a point of contention, particularly with the Trump administration. The company has expressed concerns over how the military might utilize its AI models, arguing that the systems could be overpowered or misaligned with ethical standards. In response, the administration has taken steps to limit the use of Anthropic’s latest models, Fable 5 and Mythos 5, by issuing a directive requiring the company to block foreign nationals from accessing them. This move comes as part of a broader strategy to mitigate risks associated with AI technologies before their public release.
Industry Pushback Against the Directive
More than 100 cybersecurity experts, including leaders from Adobe and Nvidia, have written to the government to oppose the directive. They argue that restricting access to Anthropic’s models could hinder the nation’s ability to defend against increasingly sophisticated cyber threats. The letter highlights that while Mythos is “quite good” at identifying software flaws, it is not the only tool capable of this task. Many experts noted that they regularly use other foundation and open-source models for security audits and training, and they warned that removing these capabilities without justification could leave the country vulnerable.
The directive was issued just 10 days after President Donald Trump signed an executive order establishing a framework for vetting AI systems. Under the order, federal agencies would evaluate the national security risks of the most advanced AI models for up to a month before their public release. The requirement for developers to participate voluntarily has sparked debate about the balance between innovation and security. Anthropic, which released a limited version of its Mythos model—Fable—this month, complied with the directive by temporarily disabling access to the latest versions. However, the company maintains that the government’s actions may not be fully justified by the risks it has highlighted.
The incident has also intensified tensions between Anthropic and the Trump administration. While the company has worked closely with US intelligence to improve its models, it has pushed back against policies it views as overly restrictive. Critics argue that the directive could stifle progress in AI development, while supporters claim it is necessary to safeguard against potential breaches. The debate reflects a broader discussion about how to regulate emerging technologies without stifling their benefits to society.
Implications for the Future of AI Security
As AI systems become more integrated into national infrastructure, the need for robust security measures grows. The Mythos test revealed that even highly sensitive systems may not be as secure as previously assumed. This has prompted calls for a more proactive approach to AI governance, including increased transparency and collaboration between tech companies and government entities. Analysts suggest that the incident could lead to a reevaluation of how AI is tested and deployed, particularly in critical sectors like defense and finance.
Experts also point to the rapid advancements in adversarial AI as a key factor in the urgency of the situation. With nations investing heavily in AI-driven cyber capabilities, the potential for misuse by hostile actors is a pressing concern. The Mythos model’s ability to detect flaws quickly highlights both its power and its vulnerability to exploitation. While the test was conducted under controlled conditions, it serves as a warning about the challenges of securing systems against AI-enabled attacks in the future.
Despite these challenges, the collaboration between Anthropic and US agencies has been seen as a positive step in understanding and mitigating AI risks. The company’s willingness to work with the government has been praised, though its concerns about the military’s use of AI remain unresolved. As the debate over AI regulation continues, the outcome of this situation may shape how the US approaches the integration of advanced technologies into its security framework. Whether the directive will lead to greater control or hinder innovation remains to be seen, but one thing is clear: the speed at which Mythos identified vulnerabilities has forced a reckoning with the security landscape in the digital age.
Industry leaders have stressed the importance of maintaining a flexible approach to AI regulation. They argue that blocking access to models like Mythos could create a vacuum in cybersecurity, allowing adversaries to fill the gap with less advanced tools. By contrast, the government’s directive aims to ensure that the most powerful AI systems are scrutinized before they are released to the public. The challenge lies in striking the right balance between preventing potential threats and fostering technological progress.
As the discussion around AI security evolves, the case of Mythos serves as a pivotal moment in the history of AI development. The model’s performance has not only demonstrated its capabilities but also highlighted the need for continuous adaptation in cybersecurity strategies. Whether Anthropic’s efforts will lead to stronger protections or increased scrutiny remains a topic of intense debate, with implications extending far beyond the immediate test results.
“This tool broke into almost all of our classified systems, not in weeks but in hours,” said Senator Mark Warner, attributing the findings to General Joshua Rudd of the NSA and US Cyber Command.
The incident underscores the dual-edged nature of AI: it can be a powerful asset in identifying threats, yet also a formidable tool for exploiting them. As the US government moves forward with its vetting framework, the role of companies like Anthropic in shaping the future of AI security will be closely watched. The outcome of this process could determine whether the nation is better prepared to defend its systems or if it risks falling behind in the race to harness AI’s full potential.
