How dangerous is Mythos, Anthropic’s new AI model?
Anthropic’s research director Dario Amodei warned on April 7 that the company’s new AI model, Mythos, is too powerful to release widely due to its ability to uncover and exploit software vulnerabilities across major systems. The lab claims Mythos has already identified severe flaws in every major operating system and web browser, including a 27-year-old undetected vulnerability, prompting a cautious approach and a new initiative, Project Glasswing, to bolster cyber-defenses before wider deployment.
Anthropic, a leading AI research lab and rival to OpenAI, has paused the release of its newest model, Mythos, citing unprecedented risks. Dario Amodei, now Anthropic’s head, echoed his past concerns from 2019 when OpenAI delayed GPT-2’s release, arguing that Mythos’s capabilities are ‘substantially beyond’ any previously trained model. The lab highlights Mythos’s ability to detect and exploit software vulnerabilities in every major operating system and web browser, including a flaw that had evaded discovery for 27 years. To mitigate these risks, Anthropic launched Project Glasswing on April 7, a program allowing companies to use Mythos to strengthen cyber-defenses before its public release. Participants include Apple, the Linux Foundation, CrowdStrike, and Google, signaling confidence in the threat’s severity. Anthropic will cover the first $100 million of costs but plans to charge five times more for Mythos access than its predecessor, Opus, raising questions about commercial incentives. The initiative reflects a broader industry shift toward cautious AI deployment, though concerns persist about open-source labs—particularly in China—prioritizing speed over safety. Meanwhile, Project Glasswing could inadvertently weaken America’s cyber-offensive capabilities by exposing vulnerabilities earlier, potentially frustrating government efforts to stockpile ‘zero-day’ exploits. Anthropic’s revenue surged to $30 billion annually from $9 billion last year, underscoring its competitive momentum. Yet the lab’s warnings carry weight due to Mythos’s demonstrated impact, including its ability to act as both a defender and an exploiter of software flaws. The pause and Project Glasswing suggest a recognition that AI advancements now outpace traditional risk-management strategies.
This content was automatically generated and/or translated by AI. It may contain inaccuracies. Please refer to the original sources for verification.