← Back to feed

Eight days after the US Commerce Department ordered Anthropic to suspend foreign nationals’ access to Mythos 5 and Fable 5, President Trump told Axios he no longer views Anthropic as a national security threat. “Not now, but a week ago, maybe,” Trump said when asked directly whether he considered the company a threat. The reversal came after meeting CEO Dario Amodei at the G7 summit in Canada.

The White House and Anthropic are now working toward a formal technical framework that standardizes how jailbreak severity gets measured, creating a defined path to restore broader access without requiring what no model can provide: zero vulnerability.

The Framework Taking Shape

The proposed system, first reported by Politico and corroborated by senior administration officials, would assess each jailbreak incident on four dimensions:

• Bypass depth: how completely the safeguard was circumvented
• Capability exposure: which previously restricted behaviors became accessible
• Attack repeatability: whether the exploit can be reliably reproduced at scale
• Operational consequences: what real-world harm could plausibly follow from the breach

The framework acknowledges the structural reality AI safety researchers have documented for years: frontier models cannot be made completely jailbreak-immune. Red-team data from June showed that even after Anthropic’s hardening efforts, Fable 5 produced confirmed harmful completions under sustained automated attack. It outperformed Opus 4.8 under identical conditions, but neither was invulnerable.

Shifting from a binary immune/not-immune test to a severity score creates a regulatory surface that can actually function. A model that resists 99% of automated attempts across all severity tiers carries fundamentally different risk than one that fails at first contact. The framework would capture that distinction and tie enforcement thresholds to it.

Timeline

1. June 12: Commerce Department export order prohibits foreign nationals from accessing Mythos 5 and Fable 5. Anthropic suspends all access globally rather than implement country-by-country filtering.
2. June 13-15: Senior Anthropic technical staff travel to Washington for direct meetings with administration officials, including Commerce Secretary Howard Lutnick.
3. June 17: Trump and Amodei meet at G7 in Canada.
4. June 19: Trump tells Axios he no longer considers Anthropic a national security threat. Politico reports the joint framework is under active development.

Anthropic’s proposal to Lutnick includes closer pre-release cooperation with the White House on future model evaluations. That would give the government visibility into model capabilities before launch, a significant structural concession that addresses the root of the concern: the export order was triggered by a jailbreak discovered post-launch, with no prior government assessment of the model’s full capability envelope.

What This Changes

If the framework reaches agreement, it marks the first time a government and an AI lab have formally agreed on technical methodology for assessing model security. The current state creates legal and commercial uncertainty for every frontier lab: restrictions can be imposed following any jailbreak, with no defined standard for when they would be lifted.

A scored, benchmarked approach makes model governance predictable. A lab could demonstrate, using agreed methodology, that a newly discovered exploit falls below the threshold warranting regulatory action. That insulates markets, enterprise customers, and international research partners from disruptions like the one that cut off 200 institutions from Mythos 5 with zero warning.

Access for Glasswing partners and verified federal agency users remains live. Timeline for broader restoration has not been announced.