Interview
The new AI model that’s alarming Washington | The Economist
- Anthropic has launched Mythos, a new AI model characterized as an extraordinarily competent cybersecurity engineer capable of executing sophisticated cyberattacks.
- Performance metrics show a significant divergence from Anthropic's previous top-tier model, Opus 4.6:
- Opus 4.6 produced two working exploits out of hundreds of attempts on a specific Firefox vulnerability.
- Mythos achieved 181 successful exploits and constructed 27 working attack chains affecting the Windows registry.
- This represents a qualitative leap rather than a marginal improvement on existing benchmarks.
- The primary risk identified is the democratization of top-tier hacking capabilities, allowing users with no technical background to deploy advanced attack vectors.
- The announcement has triggered immediate alarm among US government officials, tech leaders, and business executives, with many citing Mythos within the first five minutes of conversations.
- Secretary of the Treasury Scott Bessent and Federal Reserve Chair Jay Powell convened an emergency meeting with bank leaders to address the threat posed by the model.
- In response, Anthropic has kept Mythos inaccessible to the general public, limiting usage to the company and 11 hand-picked partners, including Apple, Microsoft, and the Linux Foundation.
- The restriction strategy mirrors Anthropic's 2019 decision to withhold GPT-2, though current leadership views the containment as necessary to prevent the leakage of dangerous capabilities.
- The intended goal is for partner organizations to utilize Mythos to identify and patch vulnerabilities in their products before the technology becomes widely available.
- Anthropic CEO Dario Amodei, long known for prioritizing safety and advocating for regulation, faces skepticism from competitors who view the narrative as marketing; however, senior personnel at partner companies corroborate the model's unique power.
- The Trump administration, historically characterized by an "accelerationist" and hands-off approach, appears to be shifting toward intervention due to the immediate severity of the Mythos threat.
- Officials anticipate a rapid transition to informal actions where the most powerful models are released exclusively to a small set of trusted companies.
- Future governance is expected to evolve into an industry-led certification framework where government and major AI builders jointly determine the safe commercialization of models.
- Policy makers are navigating a high-stakes trade-off between imposing heavy regulation (which risks US competitive decline) and inaction (which risks an AI disaster).
- There is a recognition that an "AI accident" may have been narrowly averted by the wake-up call of Mythos, prompting a reconsideration of the laissez-faire stance.
- Preparations are underway for a potential summit between US President Trump and the leader of the G7 (referred to as President G) in May to discuss global AI coordination.
- There is a growing consensus that no single nation should possess the capability to dismantle critical infrastructure, driving interest in international standards despite an environment of mutual mistrust.
- Previous tensions between the US government and Anthropic, including the classification of the company as a supply chain risk for the Pentagon, will require significant cooperation to manage the response effectively.