Here's the sequence of events, read all together, because individually each part sounds almost reasonable. Anthropic, a company founded explicitly around the belief that AI could be catastrophically dangerous, spent months before launch calling its most powerful model too risky to release to the general public. Then it released it. Then, on 13 June 2026, the US government ordered it to pull the model entirely, for all users, globally, citing national security concerns. Anthropic complied within hours.
It's a strange story. Not because any one part of it is implausible, but because all of it happened so fast.
What is Fable 5, and where does it fit?
Anthropic's main AI assistant is Claude. The naming convention they use for their most powerful models has evolved over time, and in early 2026 Anthropic had two versions of their most capable model: Mythos 5, which was the restricted preview version available to enterprise clients and researchers, and Fable 5, the public release version.
Before either of these launched publicly, Anthropic had been quite explicit in its communications that these models represented a significant capability jump. They weren't marketing. They were warnings. The company had previously described its frontier models as potentially posing catastrophic risks if deployed without sufficient safeguards. Then, in early June, they pushed Fable 5 live for general public access.
If you're reading this and wondering what AI is and what it can actually do, Fable 5 was a useful case study in the answer: apparently, quite a lot more than anyone's quite comfortable with yet.
What made the government nervous
The UK AI Security Institute, which does independent testing of frontier AI models, found in their evaluation of Fable 5 that it could exploit computer defences at a 73% success rate. That's not a marginal improvement on anything that came before. That's a capability that, if it were in a commercially available product, would represent an extraordinary jump in what any individual or group could do to another organisation's systems with minimal technical skill.
The concern wasn't purely hypothetical. The question being asked was whether foreign nationals, including state-linked actors, could access this capability through a public API and use it in ways that threaten US national security. The answer, apparently, was yes, plausibly. That was enough.
On 13 June, as reported by the BBC, Anthropic received an order to suspend access to both Fable 5 and Mythos 5. They complied by disabling access for all customers globally. Not just US accounts. Not just accounts with flags on them. Everyone.
The 73% cyber exploit rate is the number worth sitting with. Not because it's surprising that AI can do this, but because it arrived in a product anyone with a credit card could access.
The bigger fight: Anthropic vs the Pentagon
The suspension didn't happen in isolation. The Trump administration had, earlier in 2026, designated Anthropic a "supply chain risk." That's a designation historically reserved for adversarial foreign companies, Chinese telecom manufacturers and the like. Applying it to a US-founded, US-headquartered AI company was unprecedented.
Anthropic sued the Pentagon over the designation. A US federal judge ruled that the Pentagon's directive couldn't be enforced while the lawsuit proceeds, which offered some temporary protection. But that legal fight is still ongoing, and the June 13 suspension showed that the government has other levers to pull beyond the supply chain designation.
The EU, which had only just gained access to Mythos 5 earlier in June, responded to the suspension by noting that it "further underlined Europe's need for technological sovereignty." Translation: we can't build AI infrastructure on tools the US government can switch off for us without notice.
What this means if your business uses AI tools
The honest answer is: it depends on which tools you're using and how core they are to your operations.
For most South African businesses using Claude, ChatGPT, or similar tools for drafting, summarising, and general work assistance, the Fable 5 situation is mostly a reminder that these platforms are controlled by US companies operating in an increasingly complicated regulatory environment. What happened to Fable 5 could happen to any frontier model if a government decides the risk profile is too high.
If you've built operational workflows around a specific AI model, that's worth thinking about. Not in a panic way. But if your entire AI Operating System is built on a single provider's most powerful model, and that model gets pulled, you'll feel it. The mitigation is diversification where it's practical, and not building irreplaceable dependencies on any single model's unique capabilities.
For most businesses, this means using AI tools for assistance, drafting, and efficiency, things that other tools could substitute for, rather than building systems so tightly coupled to one model that there's no fallback. You can see how we approach this in our own implementations: we build for resilience, not just capability.
The practical takeaway isn't fear. It's awareness. AI tools from US providers can be suspended without warning. That's been demonstrated. Build accordingly.
The thing that's easy to miss in all the drama
The suspension, the lawsuit, the government orders: all of it is interesting. But the part that should stay with anyone paying attention is the 73% figure.
We're at a point where a commercially available AI product can exploit computer defences nearly three quarters of the time. That's a number that belongs in a classified briefing, and it was in an API you could sign up for with a credit card. The capability growth from GPT-3 to today has been well-documented, but numbers like this make it concrete in a way that abstract discussions about intelligence don't.
The speed at which these capabilities are arriving is the thing worth watching. Not any single model, not any single government decision, but the rate. Every six to twelve months, the ceiling moves. The Fable 5 story is one data point. But it's a useful one.