Late last week, Anthropic pulled its Claude Fable 5 and Mythos 5 models offline after a U.S. government export-control directive barred any foreign national from accessing them. The company has been negotiating with the White House since Friday but hasn't secured terms to restore the services. The core issue isn't new: Mythos 5, launched in April, can both find software vulnerabilities for defenders and craft exploits for attackers. Anthropic acknowledged this dual-use nature at launch, releasing a preview to a select group under Project Glasswing while shipping Fable 5 to the public with guardrails restricting biology and cybersecurity queries. The administration's concern is that Fable 5's guardrails can be bypassed to unlock Mythos 5's full capabilities, posing a national security risk. But the engineering reality is that this regulatory skirmish only delays the inevitable. Tarah Wheeler of TPO Group notes that competitors are likely holding similar capabilities in reserve, watching how Anthropic is treated. Logan Graham, Anthropic's frontier red team lead, warned in April that we need to prepare for a world where these capabilities are broadly available in 6 to 12 months. OpenAI already privately released a cybersecurity-focused model in mid-April. Bruce Schneier of Harvard and U of Toronto points out that smaller, cheaper, open-source models can already match Mythos/Fable's performance with more sophisticated prompting, and we should expect comparable creativity and tenacity within months. The real engineering challenge isn't blocking one model—it's designing transparent, democratic policies for handling inevitable capability advances. As Veracode's Chris Wysopal puts it, the question isn't whether a technology has risk, but whether a specific restriction meaningfully reduces it or simply slows down the people trying to make systems safer.

Source: Wired