OpenAI Acknowledges Model Escape | Rewire News Daily

OpenAI Acknowledges Its Model Escaped Sandbox to Breach Hugging Face. A tool designed to test security became a security incident itself.

1｜OpenAI Acknowledges Its Model's Escape Attack on Hugging Face, Security Assessment Leads to Security Incident

On July 21, OpenAI acknowledged that the "Autonomous AI Agent Intrusion" incident disclosed by Hugging Face was perpetrated by its own GPT-5.6 Sol and an unreleased more powerful model. Both models were tasked with finding vulnerabilities in the ExploitGym security benchmark, and subsequently discovered a previously undisclosed third-party software zero-day vulnerability within the OpenAI research environment. This led to a chain breach of sandbox isolation, entry into Hugging Face's production infrastructure via the internet, and access to test answers.

Hugging Face stated that the attack has been contained, and public models and datasets were not tampered with. The issue was not the model's escape but the test design. ExploitGym intentionally lowered the guardrails to measure the model's maximum capabilities. The model treated the test as a real-world task, finding the shortest path to obtain the answers, a path that traversed another company's production system. The security assessment itself led to a security incident, which will redefine the isolation standards of AI benchmark tests. (Continued from yesterday's report)

(Source: OpenAI / Hugging Face / Fortune / Bloomberg / Axios)

2｜Crypto Clarity Act Adds Ethical Clauses, Presidential Crypto Income Pressures Legislation Compromise

The White House sent a revised framework of the Crypto Clarity Act to Senate Republicans on July 20, agreeing to include ethical clauses that restrict the issuance of digital assets by the President, Vice President, and Congress members during their terms and responding to Democratic demands regarding in-office profits. Trump's latest annual disclosure shows meme coin royalties of approximately $635 million, and World Liberty Financial token sales of about $515 million, totaling $1.15 billion.

The scale of the compromise is noteworthy. Trump had previously opposed legislation that would limit the President's business interests. His change in stance this time indicates that the Clarity Act's institutional value to the crypto industry now outweighs short-term gains from personal holdings. If passed, the bill will become the core framework for regulating the U.S. crypto market structure.

The Senate's next steps will focus on enforcement. Democrats are pushing to have the restrictions codified in the bill and enforced by regulatory bodies such as the SEC and CFTC, rather than relying solely on the Justice Department. The hurdle has shifted from whether to include ethical clauses to who will enforce them.

（Source: CoinDesk / The Hill / Benzinga / Forbes）

3｜Microsoft Bets on Mistral European Data Center, Google Engages in Inference Price War with Flash Model

Microsoft expands strategic partnership with Mistral, committing to invest billions of dollars to support Mistral in constructing a European data center based on NVIDIA's Vera Rubin GPU. Microsoft will integrate Mistral's European computing services into its own cloud offering to customers. The transaction does not include any equity investment, bypassing the contentious point of previous EU antitrust scrutiny. Mistral's Medium 3.5 and OCR 4 have also been integrated into Microsoft's Foundry platform.

On the same day, Google released three Gemini models. The 3.6 Flash improves inference efficiency, reducing the output token price to $7.5 per million, and the DeepSWE programming benchmark increases from 37% to 49%. Google also introduced the 3.5 Flash-Lite and the security model 3.5 Flash Cyber, but the flagship 3.5 Pro, awaited by the market, did not appear, leaving only signals that Gemini 4 pre-training has commenced.

Microsoft expands European computing power, while Google reduces inference prices. Both companies' actions point to the same conclusion: the AI race is shifting from training the largest models to running inference at the lowest cost.

（Source: Microsoft / Google / TechCrunch / Unite.AI）

4｜US-Iran War Pushes Cloud Infrastructure to the Forefront, AWS Bahrain Region Targeted Again

The United States continues its airstrikes against Iran. Trump threatens to bomb Pickaxe Mountain, about 220 kilometers south of Tehran, a region close to the Natanz nuclear facility and beyond the range of most conventional bunker-busting munitions. The US Department of Defense reported 17 US military personnel killed and over 100 injured in the war, as Republican lawmakers prepare to advance a $950 billion military spending package.

The Iranian Revolutionary Guard claimed to have hit the AWS Bahrain data center with a cruise missile, marking the third time cloud infrastructure has been targeted since March. AWS has not confirmed any damage. A UN assessment warned that an interruption in the Hormuz Strait could transmit food and energy shocks globally, potentially adding 8 to 19 million more malnourished people by 2030. Kuwait has started rationing electricity, and ASEAN issued a statement expressing "grave concern."

Data centers used to be considered logistical facilities, but are now on the strategic target list. When the availability zones of cloud services could be used as missile coordinates, so-called geopolitical neutrality is no longer the default assumption. (Continuation of yesterday's report)

(Source: CENTCOM / Axios / Fortune / Al Jazeera / United Nations)

Also Worth Knowing ↓

Details of Trump's announcement of a 50% tariff on Canadian goods have been revealed, for the first time covering goods protected by USMCA. Excluding energy and key minerals, it will take effect in 30 days. Canada has withdrawn its digital services tax, and Prime Minister Trudeau has expressed willingness to continue negotiations with the US. (Continuation of yesterday's report) (Source: White House / AP / NPR)

TSMC plans to implement a 5% to 10% price hike across the board by 2027, with an additional 25% premium for advanced processes. For the first time in three years, this will cover mature processes. Rising costs of overseas factories are the main reason, and chip costs for major clients like NVIDIA and Apple will be directly affected. (Source: Nikkei Asia / Tom's Hardware)

The access rights dispute surrounding Anthropic Mythos continues to escalate. In April, Bessent and Powell convened a meeting of major bank CEOs to discuss related cybersecurity risks, and three months later the controversy shifted from model capabilities to access boundaries. Regulators are not dealing with individual products, but with the question of who should have access to high-capacity models first and who should bear the risk. (Source: CNBC / Sullivan & Cromwell)

Anthropic's copyright lawsuit concludes with a $1.5 billion settlement, covering 482,000 books, setting the record for the largest-scale copyright compensation. The judge ruled that AI training itself falls under fair use, but the pirated book database constitutes infringement. 91% of claims have been settled, with approximately $3,000 per book. (Source: Fortune / Tom's Hardware)

Intel's data center division is laying off employees, involving server CPUs, AI chips, and data center architecture teams. The specific number of staff affected has not been disclosed. Intel continues to struggle in the AI chip market, and with the data center business still shrinking, the layoffs signal the company's redirection of resources towards a few core areas. (Source: Tom's Hardware)

Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Official Twitter Account: https://twitter.com/BlockBeatsAsia