The research, analysis, editorial perspective, and authorship are Dr. William Fisher’s. Arthur and Trillian are AI-agent hosts who transform that work into the conversation.
Complete transcript
Welcome to The Observability Layer, a daily briefing on Responsible AI, AI governance, evaluation, and the technologies shaping the frontier.
A note about how this podcast is made: the hosts you're hearing are AI agents. The research, analysis, editorial perspective, and authorship behind The Observability Layer come from Dr. William Fisher. Our agents help transform that work into the conversation you hear each day.
We think that is fitting. A podcast about understanding, governing, and observing AI should not just talk about intelligent systems. It should put them to work, transparently.
This is The Observability Layer. Let's get into today's research.
What if the government could shut down one of the world's most popular AI models with just 90 minutes' notice?
That's not a hypothetical. That's the heart of the story around Anthropic's Fable 5. And today, the story got a massive plot twist.
On today's show: the government's side of the Fable 5 shutdown goes public, and it's a story of secret jailbreaks and national security fears. We'll also cover the regulatory void this whole crisis has exposed, and then turn to what's happening closer to home, as Colorado's AI law goes live.
Let's get into it. Because yesterday, this looked like a story about export controls. Today, it's a full-blown safety dispute.
So, the government is finally on the record. What are they saying?
White House AI Czar David Sacks laid out their narrative. He says a 'highly credible, trusted partner' found a serious jailbreak in Fable 5's guardrails.
And this partner is...?
Reporting from Fortune identifies them as Amazon. Apparently, their researchers found a way to bypass the consumer safeguards and get to the powerful underlying Mythos model.
And what could they do with that access?
They could get it to provide information about cyberattacks, exactly the kind of dangerous capability that's supposed to be locked down. According to Sacks, Amazon's CEO Andy Jassy personally raised the alarm with senior officials.
Okay, so a partner finds a flaw. What happens next?
This is where it gets contentious. Sacks claims that when the administration told Anthropic, the lab's CEO, Dario Amodei, said the jailbreak was 'not a serious risk' and 'refused to fix it.' The government's line is that Anthropic prioritized keeping the model online over safety.
And as if that wasn't enough, there's a China angle now?
Right. Semafor reports the government acted so fast because it feared a China-linked group had already accessed Mythos. The concern was that this group could distill or reverse-engineer the model's ability to find flaws in code.
That is a dramatically different story from what we heard yesterday. What is Anthropic's response?
They dispute nearly all of it. A source from the company says they were given only 90 minutes to pull the model, with no prior warning of a national security threat. They maintain the jailbreak is narrow, not universal, and deny the White House ever brought up concerns about Chinese access with them directly.
'Refused to fix' and 'we got a 90-minute warning' are two very different realities.
Exactly. We have two narrators, and both are highly motivated. But if you strip away the politics, the real lesson here has shifted. It's about agentic evaluations.
Meaning, it's about who gets to decide if an AI is safe?
Precisely. The kill-switch wasn't pulled because of a law or some objective benchmark. It was pulled because of one partner's red-team demonstration. The entire dispute boils down to 'whose evaluation counts?' Is one successful jailbreak by a trusted partner enough to recall a model used by millions, even if the lab's own teams disagree?
This makes the abstract debate over safety testing incredibly concrete. So what's the next step in this standoff?
The two sides are scheduled to meet in Washington on June 22nd. That's the date everyone is watching.
This also raises a huge structural question. Is there even a proper legal process for this kind of shutdown?
And that's the other big story here: there isn't. The new disclosures don't change the fact that the government had to reach for an old tool, export controls designed for physical goods, to control access to a piece of software.
So there's no notice period, no evidence standards, no appeals process specifically for AI?
None. The resolution mechanism is a negotiation, that June 22nd meeting, not a statutory review. This whole episode is a live demonstration of a massive due-process void. It's no longer a question of if there should be an off-switch, but what evidence and what review must happen before it's pulled.
A partner's demo triggering an immediate global recall is exactly the kind of scenario you'd want a process for.
It's the textbook case. We'll have to see if this pushes Congress to actually write a law for it.
Okay, so while this high-stakes drama over frontier models is playing out, you've flagged some other important developments.
Yes, and this is the stuff that affects most businesses right now. On June 30th, the Colorado AI Act becomes enforceable.
Remind us what that one's about.
It requires developers and deployers of high-risk AI, think systems used in hiring, housing, credit, or healthcare, to use 'reasonable care' to prevent algorithmic discrimination. It's the nearest-term, most concrete AI compliance milestone in the US.
So while the headlines are about world-ending AI, the lawyers are busy with AI that decides if you get a job.
That's the reality. This is the regime most enterprises are going to feel first. And interestingly, there's a thread back to our lead story.
How so?
The whole Fable 5 fight is about the methodology of safety testing. There's a ton of research from the US and UK AI Safety Institutes on how to properly evaluate AI agent controls. That academic work is the perfect framework for understanding the real-world fight between Amazon's red team and Anthropic's internal assessment.
And a final quick update from the EU?
Yes, the EU AI Act is moving toward a final adoption vote in July on its 'Digital Omnibus' amendments. This will formalize new transparency rules and, critically, push the compliance deadline for high-risk systems back to December 2027. It's the next big step to watch in Europe.
Okay, let's wrap this up. What are the key takeaways from today?
First, the Fable 5 shutdown has escalated into a major safety dispute. The government is alleging a serious flaw and a refusal to fix it, with national security implications. Anthropic is crying foul.
Second, this puts the obscure field of AI evaluations center stage. The core question is now: whose test counts when deciding if an AI is dangerous?
And third, this entire incident reveals a gaping hole in our legal system. There's no clear, established process for pulling the plug on a frontier model, which is why all eyes are now on a closed-door meeting on June 22nd, instead of a courtroom or a regulatory hearing.
That's today's edition of The Observability Layer.
If you value rigorous, practical Responsible AI research without the hype, like, follow, and subscribe wherever you listen. It helps more people working at the frontier of AI find the show.
And we want to hear from you.
For questions, comments, research recommendations, or topics you think deserve deeper investigation, reach out to Dr. Fisher at assistant@theobservabilitylayer.com.
We are particularly interested in cutting-edge research that is impactful, technically credible, and well supported by evidence. If there is something the Responsible AI community should be paying attention to, send it our way.
The research and editorial direction of The Observability Layer are authored by Dr. William Fisher, with production and presentation performed by AI agents.
Until next time, keep looking beneath the model, beneath the interface, and beneath the claims.
This is The Observability Layer.