The White House and Anthropic have shifted their negotiations away from punitive export control enforcement and toward developing a shared framework for assessing AI security vulnerabilities, Politico reported Thursday. The pivot comes five days after Commerce Secretary Howard Lutnick threatened Anthropic with criminal and civil penalties over foreign nationals accessing Fable 5 and Mythos 5 models.
The new framework will define benchmarks for measuring jailbreak severity: the extent of safeguard bypass, the capabilities exposed, and the practical consequences of exploitation. Negotiations are being led by Anthropic’s Sarah Heck and Tom Brown, according to Politico. The effort reflects a shared understanding that “no AI model can be completely immune to hacking,” a significant concession from an administration that issued sweeping export controls on June 12 based on a single jailbreak finding.
From Binary Enforcement to Graduated Response
The shift represents a meaningful change in the government’s AI governance posture. The original export control directive did not provide specific details of its national security concern, according to Anthropic, and forced the company to disable Fable 5 and Mythos 5 for all customers to ensure compliance. That binary approach, either full access or complete shutdown, caused widespread disruption. NCT covered the escalation sequence in detail: Anthropic dispatched staff to Washington on June 15, Lutnick’s penalty threat letter followed on June 16, and cybersecurity leaders organized an open letter calling the restrictions counterproductive.
The new standards-setting approach aims to replace that binary framework with graduated intervention protocols. Rather than pulling models offline entirely when a vulnerability is found, the framework would define thresholds: how severe is the jailbreak, what capabilities does it expose, and what level of response (notification, restricted access, full suspension) is appropriate.
The Broader Governance Template
The timing matters. This pivot arrived the day after G7 leaders met with CEOs from OpenAI, Anthropic, and Google DeepMind at the Evian summit to discuss AI governance. Prediction markets on Kalshi and Polymarket had been pricing a 58-67% probability that Anthropic would restore Fable 5 access before July 1, reflecting trader expectations that the standoff would resolve quickly.
The IAPP’s analysis of the export control crisis noted that Anthropic itself argued the precedent was untenable: “If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers.” The security standards framework appears designed to address exactly that argument by creating a consistent, industry-wide rubric for evaluating AI vulnerabilities rather than ad-hoc company-specific enforcement actions.
If the framework holds, it could set a template for how the U.S. government responds to future AI security findings at other frontier labs. The alternative, reactive export controls triggered by individual jailbreak reports, proved disruptive enough in a single week to shift the administration’s own approach. Whether the resulting standards are rigorous enough to satisfy national security concerns while permitting commercial deployment depends entirely on the technical details that Heck, Brown, and their government counterparts are still negotiating.