Recently, X's security team official account @Safety quietly dropped a "reassurance": appointing Michael O'Herlihy as the new Head of Safety, he is no longer just a traditional "content police," but is required to ensure both platform and its underlying AI systems—he must be responsible for user and brand safety, uphold the banner of "freedom of speech," and make the entire security governance process more transparent. This appointment document, for the first time, bundles "AI system security" with platform security into the hands of the same person, indicating that the narrative of social platform security is shifting from content moderation after posts are deleted or accounts are banned to systemic governance that includes algorithms, models, and infrastructure. At the same time, the South Korean Financial Commission convened major cryptocurrency exchanges and custodians for an emergency security inspection meeting, discussing raising the ratio of cold wallet assets from "over 80%" to about 99%. In contrast to Upbit's cold storage of over 99% and Bithumb's approximately 94%, this is seen by researchers as a rewriting of the industry bottom line for "user asset security" with a higher cold storage threshold; meanwhile, on another front, Anthropic is both presenting long-term plans to the media—expected to raise computing power to over 3.5GW by the end of 2026 and then double it again in the following year—and open-sourcing a zkSNARK inference verification prototype called FlashProve, which can generate about 110 proofs per second on a single H100 graphics card for the 400 billion parameter Llama 4 Maverick model, attempting to use cryptography to ensure "this is indeed what that model said," integrating verifiable AI into the so-called "trust infrastructure." From X's security chief to South Korean regulators raising the cold wallet ratio, to Anthropic using computing power and verifiable tools to endorse model credibility, these seemingly disparate actions jointly bring the issue to the forefront: In a world interwoven with regulatory clauses, platform governance, and a computing power arms race, who will define the future security boundaries?

X's New Security Chief Takes the Stage: The Clash of Freedom and Risk Control

X's first answer is to put a person front and center. The platform's security team announced via its official account @Safety that Michael O'Herlihy has been appointed as the new Head of Safety, and this time, the job description is unusually specific: he is to be responsible for user safety and brand safety on the platform, while being explicitly required to oversee the AI systems within the platform. In other words, it is not just about whether posts are deleted or accounts suspended; rather, recommendation algorithms, chatbots, and even the behaviors of the underlying models are all included in the same "safety chain." In the official statement, this chain is forced to be tied to another goal—"protecting freedom of speech." Safety, brand, AI, speech—these terms that previously belonged to different teams are compressed into one position, and the tension is almost palpable in the appointment text.

For O'Herlihy, the real challenge lies not in the title but in the gaze of the three parties: users focus on "safety," advertisers focus on "brand risk," regulators focus on "system responsibility," while some active users may use "freedom of speech" to question each intervention. X places "increasing the transparency of safety work" as a standalone responsibility, meaning that safety governance is no longer just about producing a compliance report after events or explaining the logic of post deletions at hearings, but must be prepared to clearly communicate to the outside world "why this content was pushed, and why this response was approved" at the design stage of AI systems and recommendation rules. This shift from reactive to proactive design makes the Head of Safety seem more like a "safety puzzle piece" on the map of platform and AI product expansion—only by filling this piece can X have a chance to persuade the world that it can still maintain that vaguely defined but critical security boundary while continuing to expand its global discourse influence.

Cold Wallet Hits 99%: South Korean Regulators Strike Hard

While platform safety leaders are busy rewriting internal rules, the South Korean Financial Commission (FSC) straightforwardly called major exchanges and custodians to the table for an "emergency safety inspection meeting." The briefing mentioned a core issue distilled into a figure: should the cold wallet asset storage ratio for user assets be adjusted from the current "over 80%" directly to about 99%, and if so, how? This is not an abstract discussion; the FSC referenced data from leading platforms—Upbit's cold wallet storage ratio has already exceeded 99%, and Bithumb has achieved about 94%. The numbers themselves created a silent pressure in the meeting room: since some have already run ahead, can "80% compliant" still be viewed as a safety red line?

Raising the cold wallet ratio from 80% towards "almost full offline" means exchanges must reselect their stance between operational efficiency and extreme security. As funds in hot wallets decrease, the smoothness of internal transfers, on-chain deposits and withdrawals, and large cash-outs may be squeezed, forcing a contraction in market-making and product design space. But the trade-off is a thicker firewall against hacker attacks and single points of failure. The briefing did not specify whether the meeting resulted in a final regulatory decision, but it has been classified as a "strong signal of higher requirements for user asset safety," and together with the emerging cryptographic tools, is being placed into the framework of "trust infrastructure": as global regulation tightens and technical sides attempt to prove their credibility with verifiable means, South Korea is using a cold wallet ratio approaching 99% to answer a deeper question—can this system truly be entrusted with real money.

Anthropic's Massive Computing Power and FlashProve

While regulators are raising safety thresholds with a "close to 99% cold wallet ratio," the technical answer is stacking power and transistors to unprecedented heights. According to a report by The New York Times, Anthropic disclosed to the public its mid- to long-term plan: to master over 3.5GW of computing power by the end of 2026 and plan to double that in the following year. 3.5GW is already close to the power load of a medium-sized country, and "doubling the next year" places the rhythm directly into arms race territory—not just a little more every year but exponential growth to push the training and inference capabilities of large models to a new magnitude. The research briefing did not specify where this computing power would be deployed or how much it would cost, but even looking just at the growth rate curve, it clearly illustrates a reality: in the competition of "whose model is stronger, who has more resources," computing power itself is being treated as the toughest chip.

However, Anthropic is also doing something that seems contrary to "more computing power": making model outputs verifiable. It has open-sourced a zkSNARK inference verification prototype named FlashProve, aimed not at creating a larger model but at proving that a specific response indeed comes from certain model weights without disclosing those weights, using cryptographic means to answer "did that statement really come from this model?" In tests with the 400 billion parameter Llama 4 Maverick model, a single H100 graphics card can generate proofs at a rate of about 110 per second, which is regarded by the briefing as representative of "verifiable AI" under current hardware conditions: far from zero cost, but already able to be integrated into the design space of real systems. The expansion of computing power provides raw capability to the model, while tools like FlashProve attempt to wrap that capability in a verifiable shell—one end being a power pool of over 3.5GW, preparing to double, and the other end being verifiable proof machinery like zkSNARK from the cryptographic world. The overlap means that the intersection of AI and the encryption industry is shifting from speculative narratives to issues at the infrastructure level: in the future, we must not only ask "how strong is the model," but also "who can on-chain, in regulatory scrutiny and commercial contracts, provide a proven source of the model that everyone trusts."

From X to South Korea: The Race Between Security and Computing Power

From X's newly appointed Head of Safety, to South Korea's FSC discussing raising the cold wallet ratio from "over 80%" to about 99%, and Anthropic both planning to raise computing power to over 3.5GW by the end of 2026 while open-sourcing FlashProve, a zkSNARK inference verification prototype, they all point to the same issue: how to rebuild and solidify trust in a rapidly expanding environment of business, finance, and computing power. The three entities are taking completely different paths: X chooses to make structural changes by putting user safety, brand safety, freedom of speech, and transparency within the platform and AI systems into the hands of a single role, Michael O'Herlihy; South Korean regulators align cold wallet "80% to about 99%" hard indicators with leading samples like Upbit (>99%) and Bithumb (about 94%); Anthropic uses cryptographic tools to support verification, achieving about 110 proofs per second on a single H100 for the 400 billion parameter Llama 4 Maverick, allowing for model provenance to be verifiable without disclosing weights. For participants in the encryption and AI fields, "safety," "transparency," and "verifiability" are transitioning from passive compliance costs into competitive advantages and bargaining chips—those who can quickly gain the trust of regulators, institutions, and users will have greater pricing and narrative space. What remains to be observed is whether more platforms will continue to strengthen safety roles aimed at AI systems like X, whether Korean-style technical requirements will be written into more detailed regulatory clauses, and whether verifiable tools like FlashProve can be standardized and productized to become industry's common "trust infrastructure," while short-term fluctuations in US stock contracts and oil prices are likely to continue as macro background noise in this race between security and computing power.

Join our community, let’s discuss and grow stronger together!
AiCoin exclusive Hyperliquid benefits: https://app.hyperliquid.xyz/join/AICOIN88
AiCoin exclusive Aster benefits: https://www.asterdex.com/zh-CN/referral/9C50e2
On-chain Telegram community: https://t.me/AiCoinWhaleData
On-chain community: https://www.aicoin.com/link/chat?cid=N6OVMor5g
AiCoin on-chain Twitter: https://x.com/aicoinwhaledata

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到[email protected]，本平台相关工作人员将会进行核查。