The Pentagon blacklisted Anthropic for eighteen days — and the whole time, Chinese firms actually on the blacklist were getting served just fine by somebody else. If you're just joining us, this fight started when Commerce ordered Anthropic to cut foreign-national access to Fable 5 and Mythos 5 over a contested cyber-risk claim. Access came back in stages — Fable with a safety fix attached, Mythos reportedly cleared for more than a hundred U.S. companies and agencies. Since then, we've been watching whether government-gated releases become the template for national-security AI. This is Anthropic Pentagon Watch — and today, the loophole is sitting in plain sight, with two named labs and the phrase 'completely legal' right there in the source. Let's start with the FT's reporting on who the Pentagon's blacklist was actually protecting us from. From Awesome Agents:
Existing US regulations don't bar Chinese-headquartered companies from accessing advanced AI models when they operate through subsidiaries outside mainland China. A company on the Pentagon's 1260H list can add an affiliate in Singapore, open an API account, and legally receive frontier model outputs - including from GPT-5.6 and Google's Gemini suite - every day.
So here's the irony that kind of writes itself. The week started with Anthropic as the lone blacklisted American lab. Now the Financial Times reports that OpenAI and Google were serving Singapore-registered affiliates of Alibaba, Baidu, and Tencent — every one of them on the Pentagon's 1260H list — the entire time. And the phrase from the FT is 'completely legal.' Completely legal! Anthropic gets a global shutdown for saying no to autonomous weapons, and OpenAI's out here selling to PLA-linked firms with a clean conscience and a Singapore mailing address. The mechanism matters here. The 1260H list is a DoD naming register — it doesn't carry the automatic contract prohibitions the Entity List does. So the fence that's supposed to mark the hard line on PLA-tied firms is basically paper. Right, and the enforcement gap cuts the other way too. Anthropic already blocks every Chinese company and overseas affiliate — walking away from hundreds of millions in revenue to hold that line. It got designated anyway. And the DoD's silence on OpenAI and Google? Deafening. Here's Brian Turcotte at Kilo Blog:
When Claude Fable 5 launched on June 9, we ran it through our benchmark suite, KiloBench, the same day. Three days later, the US government ordered Anthropic to take it offline, a controversial decision. When it came back on July 1 with a retrained safety classifier, we ran the exact same suite again on July 4.
So somebody actually ran the benchmark. Kilo took Fable 5 before the government-ordered safety fix and after, then counted how often you actually get the full-strength model versus the patched one. Which is the evidence version of the question everyone was guessing at. The order mandated a fix; nobody outside the lab could see what that fix cost users. And here's what gets me — we just came off the FT story where OpenAI and Google were legally selling to Singapore affiliates of 1260H-listed firms. So one lab gets its model degraded by federal order and metered by some routing table, while the other labs keep selling through affiliates with zero friction. Kilo's useful here because it's an outside party measuring degradation instead of taking a lab's compliance statement at face value. That's the accountability layer the order itself never built in. Right. The government said 'fix it,' the lab said 'fixed,' and it took a third-party dev shop to publish how often 'fixed' means 'you get the lesser model.' Transparency by benchmark, because the paperwork sure wasn't going to tell you. This one's from SC Media UK:
Yet even in during the short period Mythos and Fable were disabled, the move caused widespread disruption. On 23 June, legal AI start up Legion filed proceedings, arguing that the restriction unlawfully disrupted its business, because members of its software development team based in Canada lost access to the models.
We know the June order. The new wrinkle is who's reading it now. SC Media UK says UK security leaders are treating the Fable 5 and Mythos 5 controls as a real constraint on model access, not a U.S.-domestic squabble. So now allied security agencies are writing contingency plans around whether they can trust U.S. AI access. The export order created that governance dependency — foreign nationals inside American companies get cut off, and now Whitehall's rethinking its whole posture. And Marks & Clerk's Fiona Phillips draws a sharp line — traditional export controls on semiconductors publish restrictions in advance so businesses can structure around them. This regime didn't. You woke up with the fence already up. Which sits badly with the FT piece we just hit — a French engineer in a New York office gets locked out, but OpenAI and Google were legally serving Singapore affiliates of PLA-listed firms the whole time. UK planners are reacting to the side of the rule that hits allies, while the loophole stays open. If Anthropic Pentagon Watch helps you stay on top of the latest, take a moment to subscribe and leave a review wherever you're listening. It helps other people find the show, and it really makes a difference.
You'll find links to every story we mentioned today in the show notes. If something caught your ear, they're there for a deeper read.
That's Anthropic Pentagon Watch for today. This is a Lantern Podcast.