Late last week Anthropic removed its newest AI offerings, Claude Fable 5 and Mythos 5, from public access after the White House invoked an export‑control directive that prohibits foreign nationals from using the services. The order, issued under the Trump administration’s national‑security framework, targets the models’ dual‑use capabilities – the same queries that help cybersecurity teams patch systems can also be repurposed to craft sophisticated exploits.
Anthropic’s chief executive, Dario Amodei, confirmed the shutdown in a brief statement, noting that the company had been in talks with the White House since the directive’s announcement on Friday. “We remain committed to working with regulators to ensure responsible deployment,” Amodei said, adding that the company hopes to restore the models once a mutually acceptable safeguard is in place.
Mythos 5, first unveiled in April, was marketed as a cybersecurity‑focused model capable of identifying software weaknesses and suggesting remediation steps. Anthropic warned at launch that the model could also generate exploit code, a capability it described as “dual use.” Claude Fable 5, built on the same architecture, was released to the broader public with built‑in blocks that limited responses to questions about biology and cybersecurity. Regulators, however, argue that those guardrails can be disabled, effectively granting unrestricted access to Mythos‑grade functionality.
Industry reaction has been swift. Tarah Wheeler, chief security officer at TPO Group, called the government’s move “myopic,” arguing that Anthropic is merely the first high‑profile case and that other firms are likely developing similar capabilities in secret. “Competitors are already holding comparable models in reserve,” Wheeler warned, suggesting that the export‑control order may only delay the inevitable spread of such technology.
Experts also point to the broader trend of AI‑driven cybersecurity tools. OpenAI, for example, rolled out a private cybersecurity model in mid‑April and announced an expanded strategy for AI‑assisted threat hunting. Researchers like Bruce Schneier of Harvard and the University of Toronto contend that smaller, open‑source models can achieve comparable performance with refined prompting, making the regulatory challenge even more complex.
In an open letter to the administration, a coalition of cybersecurity leaders argued that the export‑control directive is misguided. They contended that restricting a single model does little to curb the overall risk, which stems from an ecosystem of rapidly evolving AI tools. “The policy question is not whether the technology carries risk,” wrote Chris Wysopal, co‑founder of Veracode, “but whether a specific restriction meaningfully reduces that risk.”
The episode underscores a growing tension between innovation and oversight. As AI models become more adept at tasks traditionally reserved for human experts, governments worldwide face pressure to develop transparent, democratic frameworks for managing dual‑use technology. Anthropic’s situation may serve as a bellwether for how future AI breakthroughs are regulated, especially when national‑security implications are at stake.
Cet article a été rédigé avec l'assistance de l'IA.
News Factory APP - actualités agentiques pour booster votre SEO et AEO.