Anthropic has chosen to restrict public access to its advanced AI model, ‘Mythos,’ citing large safety concerns. This decision has sparked a wide-ranging debate about the responsible governance of powerful artificial intelligence. The core tension lies between the desire to foster innovation and the need to protect society from potential misuse or unintended consequences of rapidly evolving AI abilities.
The Reported Capabilities of Mythos
Anthropic’s internal testing of Mythos reportedly uncovered a startling ability to identify critical software vulnerabilities. The company described Mythos as a “zero-day vending machine,” capable of discovering previously unknown flaws that could be exploited by malicious actors. These findings are central to Anthropic’s claims of danger.
For instance, Mythos reportedly found a 16-year-old vulnerability in FFmpeg, a widely used multimedia framework. This flaw could allow an attacker to create a malicious video file. Such a file could trick a decoder into writing a few bytes of data outside its designated memory. This action could crash a program or corrupt nearby data. The model also uncovered a 27-year-old bug in OpenBSD, a security-focused operating system. This particular vulnerability could allow a remote attacker to trigger a null pointer write, causing any OpenBSD machine reachable over TCP to crash instantly.
Beyond these specific examples, Mythos reportedly exploited several JavaScript engine bugs across major web browsers. These exploits could allow a malicious webpage to escape the browser’s sandbox environment. In one instance, this led to data theft across websites. In another, it enabled direct writing to the operating system’s kernel, giving an attacker full control over a device once a victim opened the webpage. Perhaps most concerning, Mythos found a bug in the Linux kernel. This flaw allowed it to flip a single bit in a neighboring memory page. This action turned the password executable into a writable file, which Mythos then overwrote to gain full root access to the system. The reported discovery rate of these bugs was large, with one internal tester noting they found more bugs in a few weeks than in their entire life combined.
Anthropic’s Approach to Risk Mitigation
In response to these findings, Anthropic announced ‘Project Glasswing.’ This initiative aims to secure critical global software by granting select entities access to Mythos. The idea is that Mythos is too dangerous for general public release. Instead, a collection of companies, described as those that “happen to pay Anthropic a lot of money,” along with a major bank, will gain access. The US Treasury Secretary and the Federal Reserve Chair reportedly held an urgent meeting with bank CEOs to discuss the security dangers posed by Mythos, underscoring the perceived threat level.
Anthropic’s plan is for this “fellowship” of privileged users to quickly patch vulnerabilities in the world’s software. This would happen before other organizations can build models with similar abilities. This strategy reflects a belief that controlled access is the best way to manage the immediate risks posed by such a powerful AI. It also suggests a race against time to secure systems before advanced AI abilities become more widespread.
Skepticism and Counterarguments
Not everyone is convinced by Anthropic’s claims or its chosen mitigation strategy. Some critics suggest that the company is following a familiar playbook for AI model releases. This playbook involves generating fear about a model’s abilities, only to later release a less impactful version. They point to past instances where AI models were initially hyped as revolutionary, only to have their impact diminish over time.
Skeptics also highlight several points that question the true extent of Mythos’s unique power. Anthropic has been using Mythos internally since February 24th. During this period, the company reportedly experienced its own security issues, including leaked Claude source code and documents revealing Mythos’s existence. They also faced challenges keeping their APIs online. These internal struggles raise questions about the company’s ability to manage a model deemed so dangerous.
The methods used to discover some vulnerabilities also draw scrutiny. For example, the OpenBSD vulnerability was found using 1000 parallel agent runs across the codebase. This process cost nearly $20,000 in compute resources. Critics argue that similar processes, if applied to other advanced models like Opus 4.6 or GPT 5.4 Pro, might yield comparable results. And, a claim that Mythos had an 84% success rate at writing working exploits in Firefox, a large jump over Opus 4.6’s 15%, comes with an important caveat. This success rate was achieved against a SpiderMonkey shell with process sandboxing and other mitigations turned off. Testing under such conditions does not reflect real-world browser environments. These details suggest that while Mythos may represent a step up in abilities from previous models, its perceived danger might be exaggerated or context-dependent.
The Broader Debate on AI Governance
The controversy surrounding Mythos brings the AI safety versus open access debate into sharp focus. Proponents of restricted access argue that models with the potential for widespread harm, like Mythos, must be carefully controlled. They believe that the risks to public safety, national security, and economic stability outweigh the benefits of broad availability. The fear is that such models could be weaponized or misused by malicious actors, leading to catastrophic outcomes.
Conversely, advocates for open access argue that restricting powerful AI models to a select few creates an elite “club.” They contend that this approach centralizes power and knowledge, potentially hindering innovation and equitable development. Open access allows a wider community of researchers, developers, and ethicists to scrutinize, improve, and find beneficial applications for AI. It also enables a broader understanding of AI’s limitations and risks, rather than relying solely on the assessments of its creators. The debate also touches on who should determine acceptable risk levels and how to balance the broad utility of powerful models with potential misuse scenarios.
In the end, whether Mythos will truly revolutionize cybersecurity or simply represent an incremental advance remains to be seen. The expert opinion suggests it “almost certainly will not” destroy the world, but it is “probably yes” a real step up from Anthropic’s current flagship model, Opus 4.6. For now, the public must largely take Anthropic’s word for its abilities, as access remains highly restricted.
