Anthropic's Latest AI Models Blocked, But Experts Say the Genie Won't Stay in the Bottle

Anthropic pulled its Claude Fable 5 and Mythos 5 models offline last week after a U.S. government directive barred foreign nationals from accessing the services. The company is in talks with the White House but hasn't yet reached a deal to bring them back.

Mythos 5, which debuted in April, can both find software vulnerabilities to help defenders patch them and identify ways to exploit them. Anthropic acknowledged this dual-use nature, noting that the same queries that benefit cybersecurity pros could be dangerous in malicious hands.

Initially, Mythos 5 was released only to a select group within Project Glasswing. Claude Fable 5, a Mythos-grade model, went public with blocks on biology and cybersecurity questions. But the administration moved to restrict both, believing Fable's guardrails can be bypassed to unlock Mythos's full capabilities.

Experts argue this clash misses the bigger picture. Tarah Wheeler of TPO Group says it's "myopic in the extreme" to think competitors haven't already developed similar capabilities. Anthropic's own frontier red team lead, Logan Graham, warned in April: "We need to prepare now for a world where these capabilities are broadly available in 6, 12, 24 months."

OpenAI already released a cybersecurity-focused model in mid-April. Harvard's Bruce Schneier notes that smaller, cheaper open-source models can already match Mythos's performance with more sophisticated prompting, and other models will catch up within months.

The real policy challenge, says Veracode cofounder Chris Wysopal, isn't whether a technology carries risk. "The question is whether a specific restriction meaningfully reduces that risk or whether it mainly slows down the people trying to make systems safer."

Source: Wired