Commerce Restored Fable 5 After an 18-Day Blackout

Commerce Secretary Howard Lutnick withdrew export controls on Claude Fable 5 and Mythos 5 on June 30 — ending an 18-day global blackout that began when Amazon researchers flagged a jailbreak and Anthropic had no way to nationality-gate access in real time.

What changed Tuesday: The ban is gone. The framework isn't.

• Lutnick posted on X that BIS "evaluation of the diversion risks" meant "the controls in the June 12 letter are withdrawn" — no license required for export, reexport, or deemed export of either model.
• In a letter to Anthropic chief compute officer Tom Brown, Lutnick said Anthropic agreed to "proactively detect and address security risks," collaborate on future releases, and report "malicious activity." Commerce reserved the right to reimpose controls.
• Fable 5 returns globally July 1 on Claude.ai, Claude Code, and cloud platforms. Pro, Max, Team, and select enterprise plans get Fable for up to 50% of weekly usage through July 7.
• Mythos 5 stays on a narrower Glasswing track — partial access restored June 26 for roughly 100 vetted defenders; broader rollout still coordinated with Washington.

What triggered the blackout: A jailbreak that wasn't unique.

Anthropic's June 30 blog post says Amazon researchers found a method to bypass Fable's safeguards — prompting the model to identify software vulnerabilities and, in one case, demonstrate an exploit.

Anthropic's own testing found Opus 4.8, GPT-5.5, and China's open-weight Kimi K2.7 could identify the same flaws. Every model tested could reproduce the exploit demonstration.

The company called it "routine defensive cybersecurity work" — a borderline case, not a Mythos-level offensive breakthrough. But on June 12, Commerce ordered Anthropic to cut off all foreign nationals, including its own employees. With no real-time nationality verification, Anthropic suspended both models for everyone.

The fix: Tighter classifiers, more government access.

Anthropic trained an improved safety classifier blocking the Amazon-reported technique in "over 99% of cases," per its blog. Blocked requests route to Opus 4.8. Researchers at Commerce's Center for AI Standards and Innovation tested the safeguards and called them "extraordinarily strong."

The tradeoff: more false positives on routine coding and debugging — friction users will feel while classifiers get refined.

Anthropic also pledged pre-release government testing, rapid jailbreak notification, dedicated red-team resources, and a proposed industry severity framework with Amazon, Microsoft, Google, and Glasswing partners.

The parallel nobody can ignore:

OpenAI launched GPT-5.6 Sol, Terra, and Luna June 26 under the same staggered-release model — limiting initial access to a small group of government-vetted partners while the White House reviews capabilities.

OpenAI said it "doesn't believe this kind of government access process should become the long-term default" — but accepted it anyway, calling it the fastest path to broader availability.

Tanishq Abraham, former Stability AI research director and now CEO of Sophont, told Al Jazeera: "The biggest question now is: What precedent does this set for the industry? Does the US government need to approve every frontier model release?"

Francesco Bailo of the University of Sydney's AI Trust and Governance Centre said Washington "likely realised it had overreacted" — but also that blocking Fable on these grounds would force blocking competitor models too.

What Congress still doesn't have:

Bipartisan House members' June 26 deadline for Commerce to explain the legal basis for the June 12 shutdown expired without a public answer — even as Lutnick lifted the ban four days later.

Executive Order 14409's voluntary testing framework isn't due until August. The kill switch worked. The rulebook didn't.

Convina's view: Washington didn't lose the kill switch on June 30 — it proved it works. Anthropic bought Fable back with tighter guardrails and a standing government hotline, but every frontier lab now ships on Commerce's calendar, not its own. That is not deregulation. It is pre-clearance by export control — improvised, reversible, and already the template for GPT-5.6. Silicon Valley spent two years fighting written rules and got something worse: permission slips with no statute behind them.