On July 1st, Dario Amodei testified before the U.S. Senate asking for a mandatory testing and auditing regime for the most powerful AI models. What he didn’t mention in that testimony is that Anthropic had already lived through a version of that regime weeks earlier, and not by choice.

What Amodei asked for in public, his own company already lived through by force

On June 12, 2026, Anthropic launched Fable 5 and Mythos 5, its most capable models to date. Three days later, someone found a jailbreak capable of bypassing the models’ safety guardrails, with the potential to turn them into unrestricted cyberattack tools. The U.S. Commerce Department, citing national security and export control authorities, ordered a global suspension of access to both models, including for Anthropic’s own foreign national employees.

For a bit over two weeks, the most capable model on the market simply wasn’t available to companies that had already rebuilt entire workflows around it. That blackout was a live lesson in platform risk: your stack can lose its brain overnight, and not because of anything you did.

What Anthropic accepted to get access back

On July 1st, after negotiating directly with Washington, the Commerce Department lifted the restrictions and Fable 5 became globally available again. It wasn’t free.

Anthropic committed to four things: proactively detecting and addressing security risks, helping develop shared standards for future models, rapidly reporting to the government when serious jailbreaks or malicious usage patterns are identified, and (the one that should catch the attention of any company using Claude) expanding the U.S. government’s pre-release access to evaluate frontier models and safeguards before launch, supported by Anthropic’s own technical staff.

The Commerce Department didn’t settle for a verbal promise. It explicitly reserved the right to re-evaluate those permissions and reimpose license requirements if circumstances change or Anthropic fails to deliver. That’s a mandate with teeth, not a goodwill agreement.

Anthropic also announced a new safety classifier that blocks the specific type of jailbreak that caused the suspension more than 99% of the time, and opened a HackerOne program so external researchers can report Fable 5 vulnerabilities directly. Anthropic itself warned the classifier can, in some cases, mistakenly flag legitimate coding and debugging requests. The company says most development work is unaffected, but the existence of that warning confirms something on its own: today, not even the lab that built the model can guarantee a perfect line between blocking abuse and not interfering with legitimate use.

The irony nobody mentioned that week

Here’s the part that ties it together. The same day Amodei was publicly asking the Senate for the entire industry to adopt a mandatory auditing regime, his own company was already operating under a version of exactly that regime. The difference is that Anthropic had it imposed by a cybersecurity incident, not by a voluntary corporate governance decision.

That doesn’t make Amodei’s request less legitimate. If anything, it reinforces it: the proposal isn’t hypothetical, it’s the formalization of something that already happened to the company proposing it. But it also leaves a more uncomfortable lesson for any company relying on a frontier AI vendor. If the U.S. government can shut down the most capable model on the market in three days and keep it offline for two weeks, that’s a risk variable most AI budget spreadsheets don’t account for.

What we do at IQ Source

This week I wrote about the same thing three times without planning it that way. First, that Anthropic can already audit Claude’s intent before it answers. Then, that Amodei asked the Senate to make that auditing mandatory industry-wide. And now this: the government is already exercising that power over Anthropic, with conditions that can actually be enforced. The pattern repeats across all three. The question that matters for your company was never whether the model is good. It’s how exposed you are to decisions you don’t control, made by your AI vendor or by whoever regulates your AI vendor.

In the discovery phase of AI Maestro, we map exactly this kind of platform risk before a company depends on a single vendor for critical processes: what happens if the model you chose gets suspended, how fast you could migrate a critical workflow to another vendor, and what regulatory commitments your vendor is accepting without you knowing about it. This isn’t a hypothetical exercise. Fable 5 was offline for two weeks this month, and next time it could be a model your business actually depends on.

Map your platform risk before your AI vendor decides it for you