Anthropic is broadening access to its advanced Claude Mythos AI model through an initiative called Project Glasswing. This program is designed to allow technology firms, security organizations, and government bodies to identify and rectify potential vulnerabilities and exploits within the powerful AI before its wider public release.
Key Takeaways
- Anthropic is expanding Project Glasswing, granting more entities access to its potent Claude Mythos AI model.
- A wider customer release for Mythos-class models is anticipated in the near future, following the implementation of enhanced safety measures.
- Researchers and government entities are actively evaluating the model’s capabilities in identifying security flaws and executing intricate cyber operations.
The company announced on Tuesday that it is incorporating approximately 150 new organizations into this controlled access program. This move comes as Anthropic prepares for a more extensive rollout of the model, which has garnered significant attention. The expansion follows Anthropic’s recent statement indicating an expectation to make Mythos-class models available to customers within the coming weeks, pending the completion of additional safeguards. This development coincides with Anthropic’s confidential filing for an Initial Public Offering (IPO) with the U.S. Securities and Exchange Commission on Monday.
“Following several weeks of close collaboration with our Project Glasswing partners, the security industry, open-source software maintainers, and the U.S. government, we’re extending the partnership to approximately 150 new organizations,” Anthropic stated. “Each one will need to meet our security requirements before they gain access.”
The latest phase of Project Glasswing includes organizations from crucial infrastructure sectors, such as energy, water, healthcare, communications, and hardware manufacturing. Anthropic highlighted that many of these new partners are responsible for maintaining software utilized by governments and hundreds of millions of individuals globally.
“What each partner has in common is that a successful attack on their codebase could be catastrophic,” Anthropic elaborated. “For most partners, we estimate that a major attack could affect more than 100 million people, with important ramifications for both global and national security.”
Claude Mythos first came to public notice in March following the online dissemination of draft Anthropic documents. According to the company, organizations that previously participated in the program utilized Mythos Preview to identify over 10,000 software vulnerabilities classified as high or critical severity. Partners have also leveraged the model for tasks such as generating patches, conducting security audits, and performing pre-release software testing.
The capabilities of the model have, however, raised concerns among researchers and government bodies, who fear that increasingly sophisticated AI systems could potentially lower the barrier to executing cyberattacks.
The U.K. AI Security Institute reported that during testing, Mythos autonomously executed a 32-step simulated corporate network attack. Mozilla indicated that the model detected 271 vulnerabilities that were subsequently addressed in the Firefox web browser. Additionally, the security startup Calif claimed that a preview version of the model aided researchers in developing an exploit chain targeting Apple’s M5 computer chips.
Anthropic justifies the expanded access to Mythos by asserting the necessity for defenders to possess equally advanced AI tools before malicious actors can develop comparable capabilities. The company anticipates that other developers will release similar “Mythos-class” models within the next six to twelve months, potentially without the same rigorous safety protocols. This scenario could lead to a rise in more frequent and unpredictable cyberattacks.
“In the future, frontier model releases will become increasingly high-stakes. Capabilities will continue to improve across all domains, including many that—like cybersecurity—can empower attackers and defenders alike,” Anthropic commented. “If we’re successful, we hope to enable a permanent advantage for defenders.”
Anthropic did not immediately respond to a request for further comment.
The ongoing expansion of private access to Claude Mythos has led to a decrease in confidence regarding its public release by the end of June among users on Myriad, a prediction market platform operated by Decrypt‘s parent company, Dastan. As of this writing, the probability assigned to a June public release stands at just 26.5%, a notable drop from its previous peak of 58.5% recorded last week.
Long-Term Technological Impact on Blockchain and Web3
The strategic release of advanced AI models like Claude Mythos, coupled with proactive vulnerability assessment programs, has profound implications for the future of blockchain and Web3 development. As AI capabilities grow, they present a dual-edged sword for cybersecurity within decentralized ecosystems. On one hand, advanced AI can significantly bolster the security of smart contracts and blockchain protocols by identifying complex exploits and coding errors that human oversight might miss. Project Glasswing’s methodology, which emphasizes collaborative defense against sophisticated threats, mirrors the collaborative ethos of many Web3 projects. This can lead to more robust and secure Layer 2 solutions and decentralized applications (dApps), ultimately fostering greater user trust.
On the other hand, the potential for AI to empower attackers necessitates a parallel advancement in AI-driven defensive mechanisms within blockchain. The integration of AI could lead to more sophisticated smart contract auditing tools, real-time threat detection systems for decentralized finance (DeFi) platforms, and even AI-powered decentralized autonomous organizations (DAOs) capable of self-monitoring and self-healing. The race between AI-powered offense and defense in cybersecurity will likely accelerate innovation in cryptographic techniques and consensus mechanisms, potentially driving the development of novel blockchain architectures designed with AI resilience in mind. Furthermore, as AI models become more capable, their application in analyzing on-chain data for fraud detection, regulatory compliance, and understanding user behavior within Web3 metaverses will become increasingly vital, pushing the boundaries of what is possible in data-driven blockchain innovation.
Information compiled from materials : decrypt.co