US Government Orders Anthropic to Halt All Access to Fable 5 and Mythos 5

On June 12, Anthropic announced that it had received an export control directive from the US government on national security grounds and had cut off all user access to Fable 5 and Mythos 5. The order covers all users both inside and outside the United States, including Anthropic's foreign-national employees.

Background

The government raised concerns that specific "jailbreak" techniques could bypass Fable's safeguards. According to Anthropic, however, the reported techniques are narrow and non-generalizable, limited to minor vulnerabilities that can also be found in other existing models (including OpenAI's GPT-5.5).

Anthropic says that, ahead of the Fable 5 launch, it conducted a total of thousands of hours of red-teaming with the US government, the UK AISI, and multiple private third-party organizations, and that no universal jailbreak was discovered.

Anthropic's Position

While complying with the government's legal directive, Anthropic pushed back strongly: "Withdrawing a commercial model that serves hundreds of millions of users because narrow jailbreak techniques were discovered would effectively halt new model deployments across the entire industry." The company argued that the government should have the authority to block unsafe deployments through "a legal process that is transparent, fair, and grounded in technical facts," but that this action did not follow that principle.

Other Anthropic models are unaffected, and models other than Fable 5 and Mythos 5 remain available as usual.

DiffusionGemma Is 4x Faster but Makes 6x More Fact-Checking Mistakes

A user-run benchmark has confirmed severe degradation in fact-checking for DiffusionGemma, the diffusion-based text generation model Google recently announced.

In tests run on an H100 (FP8), the standard Gemma4 26B made 5 errors out of 45 facts at 218 tok/s, while DiffusionGemma 26B recorded 28 errors out of 33 facts at 763 tok/s (roughly 3.5x faster).

Why the Errors Multiply

DiffusionGemma renders 256 tokens to the screen at once and refines them over multiple passes, but because the optimization criterion is "smoothness" alone, fabricated names, dates, and numbers survive simply because they "sound natural."

In the tests, it got the name of Steve Jobs's mother wrong, fabricated an imaginary colleague of Tetris developer Alexey Pajitnov, and listed the BeBox's price as $9,999 instead of $1,600 — errors grew more frequent the more obscure the topic.

Google itself noted at launch: "Quality is low. For use cases where facts matter, use the standard Gemma 4."

Meta Employees Sound Off on Hackathon Plan

According to WIRED, when CEO Mark Zuckerberg planned a company-wide AI hackathon, complaints poured in on an employee forum.

Some questioned whether "this company no longer supports hackathon culture," and the discussion has since grown into a broader debate over whether hackathon-style innovation can function in a large organization.

---