Another week, another multi-million-dollar hack in DeFi, and once again, it’s an off-chain compromise rather than a smart contract exploit.

AFX Trade, a decentralized perpetuals exchange that settles in dollar-pegged stablecoin USDC, was drained of about $24.15 million on Wednesday after an attacker compromised the validator signing keys behind a bridge the protocol operates on Arbitrum, blockchain data shows.

In other words, the smart contract did what it’s supposed to do – verify the signature and execute the transaction. The problem was with the private keys that generated those signatures, as attackers compromised the private validator signing keys (hot keys held offchain by the bridge operators or validators).

Steven Goldfeder, co-founder of Offchain Labs, which develops and maintains the network, said the Arbitrum native bridge “has not been hacked or exploited in any way” and that the transaction originated from a third-party protocol.

A hack of Arbitrum’s own bridge would signal risk across the entire layer-2 network, but a compromised protocol running on top of it is a contained failure.

Nothing in the bridge’s own code logic was broken. Bridges are blockchain-based tools for transferring tokens between various networks, including those they were not initially supported on.

A senior official from the White House’s Office of Science and Technology Policy has accused the Chinese AI firm behind Kimi K3 of using “covert industrial distillation” techniques to replicate capabilities from U.S. models. The allegation, posted to X on Wednesday by Michael Kratsios, underscores how U.S. concerns about AI competitiveness are increasingly blending with fears of large-scale intellectual property (IP) theft.

Kratsios said the company built an internal platform to distill U.S. models “at scale,” specifically using methods intended to evade detection. While he argued that distillation—compressing a model into a smaller one—can be legitimate and part of open innovation, he framed the alleged approach as unacceptable because it targets proprietary American technology rather than improving models through transparent research.

Key takeaways

• White House OSTP Director Michael Kratsios alleged Chinese firm Moonshot AI used large-scale covert distillation tied to the Kimi K3 release.
• Kratsios contrasted legitimate model distillation with alleged industrial-scale techniques aimed at stealing U.S. IP and avoiding detection.
• Some AI researchers dispute claims that Anthropic’s Fable was used to produce Kimi K3’s performance, citing technical plausibility and timing constraints.
• U.S. officials warned that sanctions and Entity List designations could follow IP-theft-style distillation attacks.

Why the allegation matters beyond headlines

AI distillation is not inherently controversial. In general terms, distillation helps create smaller, more efficient models by training them on outputs generated by a larger “teacher” model. The White House’s argument, as stated by Kratsios, is that scale and secrecy change the nature of the activity—turning a common engineering practice into something closer to a targeted extraction of proprietary capability.

That distinction is critical for investors, developers, and researchers because it signals a potential shift in how regulators and governments may view certain AI training pipelines. If authorities treat “covert industrial distillation” as IP theft, it could influence enforcement priorities, compliance expectations, and the willingness of model providers to share weights, outputs, or licensing terms—especially across geopolitical lines.

Timing and the dispute over Anthropic’s role

Kratsios’s claim places particular focus on the question of whether U.S. model technology was used in the preparation of Kimi K3. Cointelegraph previously reported that Anthropic’s Fable 5 was taken offline quickly due to U.S. export controls, then re-released on July 1. Kimi K3, meanwhile, launched on July 16—creating what critics describe as a narrow window for any distillation-derived transfer.

Elie Bakouch, a researcher at Prime Intellect, publicly questioned whether the technical story matches the observed outcomes. In an X post referenced in the original reporting, Bakouch argued that there are only “15 days between fable 5 ban removal and kimi K3 release,” and he added that the performance “could” not be explained in a straightforward way by distillation from Fable.

Dean Ball, head of strategic futures at OpenAI, also pushed back. On Friday, Ball said he did not believe K3’s performance could be “explained away by distillation or anything like that.” Both responses reflect a broader point: even if distillation happened, it may not be the sole—or even the primary—reason for a model’s capabilities, and establishing a clean causal link can be technically difficult.

In the absence of publicly available technical evidence, these disputes matter because they highlight uncertainty. Government accusations may have intelligence backing, but for the wider AI community, the plausibility and traceability of model-to-model influence is a separate question from whether the activity would violate policy or law.

Washington escalates from concerns to potential enforcement

The posture from U.S. officials appears aimed at deterrence. In addition to Kratsios’s claim that “covert industrial distillation” intended to steal U.S. technology is unacceptable, U.S. Treasury Secretary Scott Bessent warned that sanctions and restrictions could be pursued.

Bessent said the U.S. supports open-source AI and the innovation it enables, but he argued open source does not mean “open season” on American IP. He also warned that if firms conduct covert, industrial-scale distillation attacks that cross into IP theft, consequences could include sanctions and Entity List designations.

That statement suggests the U.S. may attempt to treat certain distillation behaviors under the same enforcement logic used for other technology-transfer and IP-protection efforts. For AI companies, the practical takeaway is that even widely used ML techniques could be reinterpreted depending on intent, transparency, and scale.

It also raises a policy tension: distillation can improve accessibility and efficiency, but enforcement actions could push industry toward more restrictive handling of model outputs and training procedures. Developers may respond by tightening documentation, auditing data provenance, or changing how they handle third-party model access.

What to watch next

Whether the dispute becomes a broader enforcement campaign will likely depend on what additional evidence, if any, is made public and how regulators define “industrial-scale” and “covert” distillation in measurable terms. For now, observers should watch for any formal government actions tied to Kimi K3 and for further clarification from researchers on what technical signals can reliably connect teacher models to student performance.

A White House official accused Moonshot AI of distilling Anthropic’s Fable AI model to develop Kimi K3, which launched last week.

In a post on X on Wednesday, White House Office of Science and Technology Policy Director Michael Kratsios alleged the Chinese AI firm developed an internal platform to distill US models at scale, using methods designed to evade detection.

“Legitimate AI distillation used to create smaller, more efficient models plays a vital role in this open innovation ecosystem,” he said. “However, large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable.”

Kimi K3 has emerged as one of China’s most capable AI models, intensifying Washington’s concerns that American models are being covertly used to accelerate China’s AI progress.

However, some AI researchers questioned claims that Anthropic’s latest AI model was used to train Kimi K3. 

Anthropic’s Fable 5 was re-released on July 1 after it was quickly taken offline due to US export controls, while Kimi K3 launched on July 16, giving a narrow window for distillation attacks to occur.

“There are only 15 days between fable 5 ban removal and kimi K3 release,” said Elie Bakouch, a researcher at AI startup Prime Intellect.

“I don’t think claiming that K3’s performance comes from fable distillation (even if they did it) makes sense technically.”

Dean Ball, OpenAI’s head of strategic futures, said on Friday he didn’t believe the K3 model’s performance could be “explained away by distillation or anything like that.”

Related: Anthropic to bring back Fable 5 as US lifts export controls

US Treasury Secretary Scott Bessent warned that the large-scale distillation attacks could result in sanctions and other restrictions.

“We support open-source AI and the innovation it unlocks. But open source is not open season on American IP,” said Bessent.

“When PRC firms conduct covert, industrial-scale distillation attacks that cross the line into IP theft, sanctions and Entity List designations will be on the table.”

Magazine: Thai scammer’s $122M wallet, Japan embraces crypto credit: Asia Express

Franklin Templeton’s head of digital assets and innovation says AI agents are poised to become a major demand driver for blockchain networks—specifically the protocols that can support rapid, low-cost payments between machines.

Speaking in a long-form post on X on Wednesday, Sandy Kaul argued that the “agentic AI” economy will require settlement speeds and fee structures that legacy card rails struggle to deliver. He pointed to blockchain ecosystems such as Aptos, Solana, and BNB Chain as better aligned with that needs-based shift.

Key takeaways

• Franklin Templeton’s Sandy Kaul links AI agents to increased demand for blockchain protocols that can handle machine-to-machine micropayments.
• Kaul argues traditional payment cards are a poor fit for agentic payments due to fees and slow settlement compared with blockchain transaction finality.
• A joint Visa and Artemis report contends card-based infrastructure is insufficient for AI agents that require near-zero fees and fast settlement.
• According to that Visa-Artemis report, the x402 payment protocol processed $15 million in adjusted volume across 109 million+ adjusted transactions since its May 2025 launch.

Why AI agents change the payment requirements

The central thesis is that agentic systems—software that can act autonomously on behalf of users or other systems—will generate a different kind of commerce than today’s human-driven transactions. Kaul framed the opportunity as an evolution beyond the way investors typically approach AI: rather than focusing only on companies “aligned” with AI, he suggested the market may also reward infrastructure designed for automated execution and continuous micro-interactions.

In his view, the payment layer becomes a bottleneck if it cannot support high-frequency, small-value transfers. Agentic micropayments are likely to be time-sensitive and cost-sensitive, meaning even modest frictions—such as higher fees or longer settlement—can make recurring machine payments economically unattractive.

Legacy cards vs. settlement speed

Kaul’s argument is not that card networks are obsolete, but that they were engineered for a different pattern of usage: relatively low-frequency human commerce where settlement delays are rarely a primary constraint.

He highlighted that visa network settlement can take one to three business days, while certain blockchain networks can finalize transactions in seconds. That timing gap is likely to matter when agents are coordinating continuously, where delays can ripple through workflows and reduce the viability of rapid settlements.

Kaul also pointed to “high fees and settlement times” as the factors that, in his assessment, make traditional payment rails unsuitable for agentic micropayments.

Visa and Artemis: infrastructure gaps for “agentic” commerce

The Franklin Templeton executive’s remarks align with a joint report released last Wednesday by Visa and investment thesis platform Artemis. In that report, the partners argue that conventional cards built for human-scale payments are not designed for the demands of AI agents.

Visa and Artemis specifically emphasize that agentic payments require infrastructure with near-zero fees and faster settlement to make micropayments commercially viable. The report’s framing reinforces Kaul’s thesis that the real battleground is payments throughput and cost efficiency—not just AI capabilities at the application layer.

Importantly for readers, this is not presented as a purely speculative concept; the report also points to existing machine-payment experimentation and early adoption signals, including activity tied to x402.

What “early adoption” looks like: x402 activity

In the Visa-Artemis report, the x402 payment protocol is highlighted as an example of a machine payment rail showing measurable usage. The report claims that x402, developed by Coinbase, processed $15 million in adjusted volume across more than 109 million adjusted transactions since its May 2025 launch.

For investors and builders, the value of that statistic is less about any single figure and more about the direction it suggests: that machine-payment protocols are beginning to attract usage under a framework designed for frequent transfers. Still, it’s also worth noting the metric is reported as “adjusted volume” and “adjusted transactions,” so readers should treat it as an operational indicator from the report rather than a direct translation into end-user revenue or broader market share.

Signals from payments providers

While the Visa-Artemis analysis criticizes card-based infrastructure as insufficient for agentic needs, the companies are also actively exploring how the broader payment ecosystem might support agentic behavior.

Visa’s crypto-related division and Stripe-backed Tempo launched AI tools in March, according to coverage referenced in the same context. Visa’s offering is described as enabling same-day payments—an attempt to address speed constraints that agentic micropayments depend on.

In parallel, Kaul’s remarks point readers to blockchain environments where settlement speed is structurally faster, suggesting a practical mismatch: even if card providers add features to move payments more quickly, the fee and settlement model may still not align with the economics of high-volume, machine-to-machine exchanges.

Going forward, the key thing to watch is whether agentic payment demand materializes in a way that drives sustained usage of low-fee, fast-settlement rails—particularly as protocols like x402 and newer infrastructure compete to serve recurring micropayment flows. The open question remains how quickly mainstream agent deployments will scale enough to make settlement and fee constraints decisive rather than theoretical.