Anthropic's Safety Superpower
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Anthropic publicly released Fable two months after deeming its predecessor Mythos Preview too dangerous for public access, citing advanced cybersecurity capabilities, then later said the U.S.-mandated suspension was a 'misunderstanding' over a non-universal jailbreak
- The U.S. government issued an export control directive at 5:21pm ET requiring Anthropic to abruptly disable Fable 5 and Mythos 5 for all customers — including foreign-national Anthropic employees — to ensure compliance, with no impact on other Anthropic models
- Amazon reportedly reported the jailbreak that triggered the directive, despite being both an Anthropic investor and a major provider of inference to the company
- White House officials have characterized the situation as 'insouciance' by Anthropic's leadership to legitimate national security concerns, while Anthropic argued the vulnerabilities found were 'relatively simple' and discoverable by other publicly-available models without any bypass
- Anthropic quietly extended Fable data retention to 30 days for all customers — including enterprise plans that had previously promised zero data retention — and put in no third-party safeguards against future training use
- SemiAnalysis estimated a $200 subscription delivers roughly $8,000 of Claude tokens and $14,000 of Codex tokens, illustrating the scale at which frontier labs subsidize access to harvest real-world usage data
- Satya Nadella of Microsoft warned that 'there is no societal permission for an AI future that hollows out entire industries,' arguing firms must build 'human capital and token capital' learning loops rather than cede value to a few dominant models
Why it matters: The export control directive crystallizes the frontier-lab bind: releasing increasingly capable models is the only path to the user-touchpoint moat and the real-world data those labs need to improve, but each release risks ceding national-security leverage to a government Anthropic has now openly clashed with. Anthropic's 30-day data retention shift, applied even to enterprise customers who had been promised zero retention, is the unstated price of staying in that race — and Amazon's dual role as jailbreak reporter, investor, and inference provider exposes how thin the walls between competitors really are.
