China Named In Frontier AI Model-Copying Advisory

CISA advisory AA26-251A, published alongside the Justice Department and intelligence-community statement this week, names six China-based AI firms running what the agencies call industrial-scale distillation against American frontier models since late 2024. The mechanism is not a hack. It is billions of tokens pulled through paid API access, then used to train competing models on the outputs. Dark Reading's framing, "covertly draining," describes a business relationship the vendors themselves billed for. The advisory names OpenAI, Anthropic, Google Gemini and SpaceX's Grok as the source models. No CVE attaches to this. The extraction ran through terms-of-service violations and output logging, not an exploit.

For a Hong Kong or Singapore bank, this lands on model-supply-chain due diligence rather than the AI governance frameworks written for internal agent use. HKMA has no circular addressing frontier-model provenance; SFC's September circulars this week cover virtual asset intermediaries and private-market fund exposure, not AI vendor risk. If a bank's AI red-teaming or vendor-risk process depends on assumptions about a foundation model's training lineage, this advisory is the first government-sourced document stating that lineage is contested at the model layer, not just the data layer.

No vendor an APAC bank runs is implicated here; Anthropic, OpenAI and Google are named as victims, not as parties requiring a patch. The control that matters is API rate-limiting and output-watermarking on the bank's own model access, the same technique used against it now available as a case study for any vendor evaluating outbound token exposure. The advisory changes what "training data provenance" means in a vendor questionnaire. It does not change anything in a SOC.

The accusation, reported by Ars Technica this week, is that six Chinese AI firms lifted outputs from US frontier models, Anthropic's and OpenAI's among them, to train their own systems rather than building from scratch. Call it what it is: distillation without a license, training a smaller model on a bigger one's answers. It's a real practice and a real dispute. But the more useful number is not how many firms got named. It's how many of them have an enterprise sales channel. As of this month, essentially none does. Alibaba Cloud and Tencent Cloud sell compute. Neither operates the kind of packaged, compliance-cleared AI product that a bank in Singapore can put in front of a loan officer next quarter. That's the gap that decides who profits from a good model, not who trained it.

Here's why that gap matters more than the lawsuit. A copied model is cheap to produce and hard to sell into a regulated institution, because the buyer isn't just buying weights, it's buying a vendor relationship: a signed contract, an integration into existing systems, and someone accountable when the Monetary Authority of Singapore asks how the model was validated. Anthropic and OpenAI have that relationship built through Azure and enterprise API contracts already live inside banks and insurers across Hong Kong and Singapore. A Chinese lab that copied GPT-6 Astra's outputs still has to build the sales team, the compliance paperwork, and the deployment track record from zero, and export controls on the H200 and equivalent hardware mean it's doing that build without the training compute the US labs use to keep improving the product it's supposedly copying. The accusation is about training data. The constraint that actually bites is that copying gets you a model, not a customer, and by the time any of these six firms builds a sales channel to match, the US labs it copied will have shipped two more model generations on hardware the copiers still can't buy.

Howitzers do not deter. They range a target and then hit it. The South China Morning Post reported this week that People's Liberation Army units have moved long-range howitzer batteries into position facing Taiwan, a deployment that follows Taiwan's navy tracking a dozen PLA warships in the waters around the island earlier this month. A howitzer's job is fire support for troops crossing terrain, not the kind of signaling a destroyer patrol can pass off as routine presence. Beijing's Ministry of National Defense, under Dong Jun since 2023, has not commented on the specific placement.

Washington's own read, delivered this week by a senior American diplomat in Taipei, put a number on what the artillery is aimed at: a war over Taiwan would produce an economic shock larger than World War II, closing a shipping corridor that carries a large share of global trade and halting semiconductor output the rest of the world depends on. That warning and the howitzer imagery describe the same theater from two directions, one counting the cost of a war nobody has started, the other positioning the hardware that starts it. Artillery ranged on a coastline is not a message meant for negotiation. It is equipment staged for the day negotiation ends.

Southeast Asia's shift on Taiwan, reported this week by Foreign Policy as regional capitals recalibrate public language on the strait, lands first on Kuala Lumpur's and Jakarta's trade desks, but the balance sheet that actually prices the risk sits at the PBOC's open-market operations desk in Beijing. That desk has been running larger seven-day reverse repo injections through early September, the standard tool for absorbing a yuan liquidity squeeze when offshore holders start trimming mainland exposure ahead of a flashpoint, and the timing sits alongside a week in which Taiwan's navy tracked a dozen PLA warships and, per the South China Morning Post, longer-range howitzer batteries moved into position facing the island, the deployment Mei Chen's brief today reads as equipment staged for after negotiation ends. The PBOC does not comment on individual repo sizing, but the desk's pattern, larger injections clustered around weeks of elevated strait activity, has held since the spring.

Or, more precisely, the desk is not defending the yuan against a single shock; it is pre-funding the offshore banks and insurers that would otherwise need emergency dollar liquidity if Taipei-Beijing tension forces a rapid unwind of mainland-linked positions, the same mechanism the finance ministry leaned on when it funded the September bank and insurer rescue partly through the state tobacco monopoly. Southeast Asia's language shift on Taiwan matters to Washington's diplomats; it matters to the PBOC's desk only insofar as it changes how fast offshore capital moves. The open-market desk has until its next scheduled reverse repo auction, typically Tuesday, to decide whether the injection size continues to track the strait's temperature or reverts to ordinary quarter-end seasonality.

The claim runs back to March 2022, when the LME suspended nickel trading and cancelled roughly $12 billion of trades in a single session after prices doubled in hours on a short squeeze tied to Tsingshan Holding Group's position. Elliott, which held nickel contracts caught in that cancellation, has pursued the LME (majority owned by HKEX since the 2012 acquisition) through the UK courts since 2022 and lost at both the High Court and Court of Appeal. The nickel trading legal claim reported this week names HKEX alongside the LME, extending a dispute that has already run more than four years without a payout to Elliott.

Look, the mechanics matter here more than the headline. HKEX bought the LME for £1.4 billion in 2012 specifically for its role as the world's base-metals price-setter (the kind of infrastructure asset that is supposed to generate steady clearing fees, not courtroom bills). A claim that reaches past the exchange itself and into its parent tests whether an owner can be held to account for a subsidiary's market-intervention decisions, which is a different and larger question than whether the 2022 cancellations were lawful. HKEX has not disclosed a provision for this litigation in its filings, and until it does, the exposure is a legal question, not a balance-sheet one.

Fujitsu says it has built a working prototype of a quantum computer whose qubits are not exotic circuits or trapped ions but flaws in diamond crystal, tin atoms sitting where a carbon atom should be, called tin-vacancy centers. Each defect traps a single electron whose spin, the axis it prefers to point along, becomes the qubit: the unit of quantum information that can hold a 0, a 1, or both until measured. Fujitsu wired the diamond defects into photonic integrated circuits, chip-scale channels that carry the light used to read and link the spins, according to the company's September 2026 announcement. The pitch is that spin qubits in diamond hold their state far longer than the superconducting loops IBM and Google use, because the defect sits shielded inside a solid crystal rather than exposed on a chip surface at near absolute zero.

The stage is benchtop prototype, one working unit demonstrated, not a machine anyone can rent time on. What diamond spin qubits are supposed to buy back is coherence time, how long a qubit holds its state before noise scrambles it, which sets the ceiling on how long a calculation can run before it needs error correction. IBM's approach on Heron chips leans on brute qubit count and lower gate error rates, matching a materials simulation this August that Japan's Fugaku supercomputer failed to reproduce after 700,000 processor-hours. Fujitsu's bet is different: fewer, longer-lived qubits linked by light instead of wires. The next gate is a published qubit count and gate fidelity number benchmarked against an existing platform, because a working prototype with no error rate on record is a proof the physics works, not yet a computer anyone can compare.

Researchers running a randomized trial in older adults with depression assigned one group to a course of cognitive behavioral therapy and compared their blood-based epigenetic clocks against a control group over the treatment period, according to the study reported this week in the journal covered by Medical Xpress. The therapy group showed a slower rate of biological aging by that marker than controls. The mechanism is not mystical: chronic depression keeps cortisol elevated and inflammatory signaling switched on, and both accelerate the DNA methylation changes that epigenetic clocks are built to detect. Reduce the psychological load that is driving the stress response, and the clock that reads out that stress response moves slower too.

What the trial establishes is narrow and still real: a specific psychological intervention, delivered by trained therapists over a defined course, changed a validated biomarker of aging in people who were clinically depressed to start with. What it does not establish is that therapy, or stress reduction generally, extends lifespan in people who were not depressed, or that the epigenetic clock used here predicts who gets sick and when, a question these clocks still answer imperfectly across populations. The honest read is that the nervous system and the aging clock are more connected than the anti-aging supplement aisle usually credits, and that the connection ran through actual mental illness, not the ordinary stress of a bad week.

Lamine Yamal broke Kylian Mbappe's Champions League goal-involvement record on Tuesday, reaching the mark faster than any player in the competition's history. He is 18. Mbappe set the standard he just erased while playing for Monaco and PSG across a run of seasons that took him until his twenties. Yamal did it before he can legally rent a car in most of Europe, on a Barcelona contract that runs through 2031 and, by the terms Barcelona structured before his stock went vertical, at a release clause of one billion euros that only looks enormous until you compare it to what Real Madrid is currently paying Mbappe to be merely excellent.

That gap is the whole Barcelona business plan in one sentence. Every La Liga club that let a generational talent walk for a knockdown fee (Barcelona itself has done it before, ask anyone who remembers Neymar's exit math) is watching this contract as the corrective. Real Madrid, Manchester City and PSG have the revenue to make any release clause look like a rounding error, but a billion euros stops being a number and starts being a deterrent. Mbappe held the record. Yamal now holds the record and the contract that makes the record someone else's problem to buy.