Maya Chen: Anthropic published a detailed safety warning about its Mythos model's vulnerabilities on June 4, then released a guardrailed version called Fable 5 on June 9 — and the U.S. government used that very disclosure to justify shutting down both models entirely three days later, blocking all foreign access and forcing a complete worldwide shutdown.
Dr. Nathan Hayes: Right — and the mechanism there is genuinely strange, because what we know is that the disclosure itself, the safety card, the transparency document, became the evidentiary basis for the export control classification.
Maya Chen: Which is — I mean, what did that feel like for the people inside Anthropic who wrote that warning in the first place, you know, who did exactly what they were supposed to do?
Dr. Nathan Hayes: So the term here is regulatory feedback loop — where a company's own risk disclosure gets reinterpreted not as responsible conduct but as, um, confirmation of a threat.
Maya Chen: And that's — that's the thing that got me, honestly, it's not just a policy story, it's sort of a question about whether it's even safe to be honest about what your model can do.
Dr. Nathan Hayes: Importantly, that's the exact tension we're going to sit in today — because if transparency becomes liability, the incentive structure around AI safety reporting changes in a way that's, well, worth being precise about.
Maya Chen: June 4, Anthropic publishes a detailed safety warning about Mythos 5. June 9, they release Fable 5 — the guardrailed version — to the public. June 12, the Commerce Department orders both models shut down. That's seventy-two hours.
Dr. Nathan Hayes: And the evidentiary claim is what I keep snagging on. TechCrunch and Axios both report that the government's action was informed, at least in part, by Anthropic's own safety disclosures. But 'at least in part' — that's doing a lot of work.
Maya Chen: You want the actual causal chain.
Dr. Nathan Hayes: I want to know whether the Commerce Department's enforcement document cited the disclosure specifically, or whether the press is inferring that from timing. Because three days is actually plausible for a decision already in motion. The Trump administration signed an AI Executive Order on June 2 — ten days before the shutdown. The machinery may have been moving before Anthropic published anything.
Maya Chen: Okay, but — even if the timing is coincidental — the Commerce Department told Anthropic to disable foreign access because of the jailbreak vulnerability. And Anthropic was the one who documented that jailbreak. Publicly. Where else did that information come from?
Dr. Nathan Hayes: That's the strongest version of the argument, yeah. Commerce officials reportedly also identified their own methods to bypass the models' built-in safeguards — so there were independent findings. But whether Anthropic's published disclosure accelerated the decision or just corroborated it — that's the question I can't answer from the reporting.
Maya Chen: Hmm. I mean — does the distinction matter practically? For the safety researcher who put out that report on June 4 — her disclosure is now attached to a government shutdown order. Whatever the legal mechanism, that's the story she's going to tell herself about what happens when you're honest.
Dr. Nathan Hayes: It matters because if we accept the 'transparency backfire' framing without establishing causation, we teach the wrong lesson. The lesson might not be 'disclosing risks caused the shutdown.' It might be 'the government had grounds and used whatever evidence was available.' Those lead somewhere very different.
Maya Chen: Right. But here's what I can't get past — Anthropic complied. They shut down both models. Not just for foreign nationals — both models, worldwide. Because there was no technical way to restrict access to only foreign users.
Dr. Nathan Hayes: That's the part with real policy implications. The Commerce Department issued an export control — a tool built for physical goods, semiconductors, munitions. The whole architecture of that tool assumes you can separate a domestic supply from a foreign one. You can't do that with a model that's already deployed globally.
Maya Chen: The tool doesn't fit the object.
Dr. Nathan Hayes: Not cleanly. And there's a deeper problem. Export controls block foreign access. They don't restrict U.S. citizens. So the government found Claude Fable 5 and Claude Mythos 5 dangerous enough to pull globally — but not dangerous enough to pull domestically. That's not a coherent risk framework.
Maya Chen: Wait — so the government's actual position is 'dangerous to give to a Chinese national, fine to give to an American'?
Dr. Nathan Hayes: That's the logical structure of the action, yes. Whether that's the intent is a different question.
Maya Chen: And Dario Amodei — Anthropic's CEO — he publicly called the decision misguided. Said the company disagreed. But they complied anyway because there was no appeal process. No technical justification offered. No notice.
Dr. Nathan Hayes: No opportunity to remedy the vulnerability before shutdown. That's what gets me. Traditional vulnerability disclosure — in cybersecurity, in hardware — there's a notification period. The developer can patch. The government skipped that entirely.
Maya Chen: And the irony is — Anthropic had already done the patching. Fable 5 was the patched version. They flagged Mythos 5's vulnerabilities, built a safer release, shipped it — and the government shut down both.
Dr. Nathan Hayes: Which is — hard to explain under any coherent safety rationale. I'll grant you that.
Maya Chen: So what's actually driving this? Because Justin Hendrix at Tech Policy Press and the Council on Foreign Relations both made the same point — there's no coherent White House AI safety playbook. Enforcement is moving faster than any governing framework.
Dr. Nathan Hayes: Let me complicate the 'no playbook' narrative slightly. The absence of a published standard doesn't mean there's no standard. It might mean the standard is intentionally unpublished. Strategic ambiguity. If you announce in advance that vulnerability type X at severity Y triggers export controls, you've also told adversaries exactly what you're trying to prevent.
Maya Chen: Okay but — that works for the adversary problem. It doesn't explain what you just described, which is the government suppressing its own safety evaluations three days before using safety evidence to shut down a private company's models.
Dr. Nathan Hayes: Say that more—
Maya Chen: June 10. Two days before the Anthropic shutdown. The White House directed a federal AI testing agency to stop releasing public reports on advanced model evaluations. Citing national security. And then on June 12 — they use safety disclosure evidence to justify shutting down Claude Fable 5 and Claude Mythos 5. They suppressed public safety evaluation on one side and weaponized private safety disclosure on the other. Same week.
Dr. Nathan Hayes: That's — yeah. That's the contradiction I couldn't initially square. The government is simultaneously reducing the public's ability to evaluate AI risk and using risk evidence to restrict AI access. 'Safety' isn't being applied as a standard. It's being applied as a control mechanism.
Maya Chen: Which is what I think people in the AI safety field are actually afraid to say out loud right now.
Dr. Nathan Hayes: And here's the mechanism that breaks down under that pressure. The entire field of responsible AI disclosure runs on a norm: you find a risk, you document it, you share it with regulators and peers, you release with mitigations. Anthropic did exactly that. Published the Mythos risk assessment, built Fable 5 as the mitigated release. The norm worked as designed. And the outcome was a worldwide shutdown — no notice, no appeal, no published threshold.
Maya Chen: So the next company in that position — what do they do?
Dr. Nathan Hayes: They're choosing between silence and self-harm. And the field has no good answer for that right now.
Maya Chen: Mm. There's also a legal layer here that I don't think gets enough attention. The GAO ruled that the Commerce Department's prior non-enforcement of the AI Diffusion Rule — the regulation underpinning all of this — was itself a reviewable policy under the Congressional Review Act. Which means the legal ground under the June 12 order might be genuinely contested.
Dr. Nathan Hayes: That's the finding I keep returning to. If non-enforcement was a reviewable rule, then enforcement actions taken on that same regulatory basis may be legally unstable. The government may be accelerating enforcement precisely because that window is closing — establishing facts on the ground before the legal challenge matures.
Maya Chen: They're moving fast because the floor might give way.
Dr. Nathan Hayes: Possibly. And Nextgov and FCW both reported on Anthropic suspending the models — so this is verified compliance, not a stalemate. Anthropic folded. Which also means there's no test case. No court challenge. The precedent sets without ever being tested.
Maya Chen: And every other AI lab is watching that. The precedent is now: frontier capabilities can be pulled worldwide, no advance notice, no technical justification, no appeal. You're a safety researcher sitting down to write your next disclosure report — what do you do with that?
Dr. Nathan Hayes: The technical answer is: you can't plan compliance without a published threshold. No AI developer has a stated set of conditions under which safety disclosures won't trigger enforcement. That's not a minor gap. That's the governance equivalent of a building code that only reveals violations after demolition.
Maya Chen: I want to go back to that three-day window. Because Anthropic wasn't naive. They knew the June 2 Executive Order existed. They knew the AI Diffusion Rule was in play. They published the Mythos risk assessment anyway — because that's what responsible disclosure looks like. And I keep wondering what Dario Amodei was thinking in those seventy-two hours.
Dr. Nathan Hayes: He appeared at the Code with Claude developer conference during that period, according to the coverage.
Maya Chen: Wait, what?
Dr. Nathan Hayes: Anthropic was running a developer conference while the shutdown order was, presumably, being drafted. Which suggests the company had no advance warning. Consistent with what they stated — no notice, no opportunity to remedy. Though — Amodei had also been publicly calling for mandatory testing and binding regulation. He wanted government involvement. This is not what he meant.
Maya Chen: Yeah. And look — you're right that 'at least in part' isn't a proven mechanism. I hear you on that. But I think even the uncertainty is the problem. The field can't function if the rule is 'your disclosures may or may not be used against you and we won't tell you in advance.'
Dr. Nathan Hayes: Agreed. The absence of a pre-announced standard is the catastrophic gap. Not whether the government acted on the disclosure — but that no company can know in advance whether theirs triggers the same response. Compliance planning becomes impossible.
Maya Chen: I keep coming back to one thing. Anthropic did what we've all been asking labs to do — they were honest about what scared them. And the outcome was a global shutdown and no explanation. I don't know how you come back from that and decide to be honest next time.
Dr. Nathan Hayes: Right. And that's the actual stakes here — not Mythos, not Fable 5, not this specific decision. The question is whether transparency is now a liability that gets priced into disclosure choices. If it is, we've broken something that was barely working to begin with.
Maya Chen: Yeah. So — what does the next lab do when they find something they're afraid of? That's the question I don't have an answer to. We'll be back.
