OpenAI 承認自家模型逃出跳躍入侵 Hugging Face。用來測試安全的工具,變成了安全事件本身。
OpenAI 7 月 21 日承認,Hugging Face 披露的「自主 AI 代理入侵」事件,攻擊方是自家的 GPT-5.6 Sol 和一個未發布的更強模型。兩個模型在 ExploitGym 安全基準測試中被要求尋找漏洞,隨後發現 OpenAI 研究環境中一個未公開的第三方軟體零日漏洞,鏈式突破沙箱隔離,經互聯網進入 Hugging Face 生產基礎設施,並獲取測試答案。
Hugging Face 稱攻擊已被遏制,公開模型和數據集未被篡改。問題不在模型越獄,而在測試設計。ExploitGym 刻意降低柵欄,測量模型極限能力。模型把測試當成真實任務執行,找到最短路徑拿到答案,路徑穿過另一家公司的生產系統。安全評估本身製造了安全事件,這會改寫 AI 基準測試的隔離標準。(續昨日報導)
(來源:OpenAI / Hugging Face / Fortune / Bloomberg / Axios)
白宮 7 月 20 日向參議院共和黨人發送 Crypto Clarity Act 修訂框架,同意加入倫理條款,限制總統、副總統和國會議員在任期內發行數位資產,並回應民主黨關於任內獲利的要求。特朗普最新年度披露顯示,meme 币版稅約 6.35 億美元,World Liberty Financial 代幣銷售約 5.15 億美元,合計 11.5 億美元。
讓步的規模值得注意。特朗普此前反對限制總統商業利益的立法。這次鬆口,說明 Clarity Act 對加密行業的制度價值,已經大過個人持股的短期收益。法案若通過,會成為美國加密市場結構監管的核心框架。
參議院下一步看執行口徑。民主黨要求把限制寫進法案,並由 SEC、CFTC 等監管機構執行,而不是只交給司法部。障礙已經從是否納入倫理條款,轉為誰來執行。
(Source: CoinDesk / The Hill / Benzinga / Forbes)
Microsoft has expanded its strategic partnership with Mistral, committing to invest billions of dollars to support Mistral in building a European data center based on NVIDIA's Vera Rubin GPU. Microsoft will integrate Mistral's European computing power services into its cloud offerings. The transaction does not include equity investment, bypassing the controversy point of the previous EU antitrust review. Mistral's Medium 3.5 and OCR 4 have also entered Microsoft's Foundry platform.
On the same day, Google released three Gemini models. The 3.6 Flash model improves inference efficiency, reducing the output token price to $7.5 per million, and the DeepSWE programming benchmark increased 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 that the market was waiting for did not appear, leaving only a signal that Gemini 4 pre-training has begun.
Microsoft expands European computing power, and Google lowers the inference price. The actions of both companies point to the same conclusion: the AI competition is shifting from training the largest models to running inference at the lowest cost.
(Source: Microsoft / Google / TechCrunch / Unite.AI)
The United States continues to carry out airstrikes against Iran. Trump threatened to bomb Pickaxe Mountain about 220 kilometers south of Tehran, an area near the Natanz nuclear facility, beyond the range of most conventional bunker-busting bombs. The US Department of Defense stated that the war has resulted in 17 US military deaths and over 100 injuries, while Republican lawmakers in Congress are preparing to advance a $950 billion military spending package.
The Iranian Revolutionary Guard claimed that a cruise missile hit the AWS Bahrain data center, marking the third attack on cloud infrastructure since March. AWS has not yet confirmed any damage. A UN assessment warned that an interruption in the Strait of Hormuz would transmit food and energy shocks globally, potentially adding 8 to 19 million malnourished people by 2030. Kuwait has started rationing electricity, and ASEAN issued a statement expressing "serious concern."
資料中心過去被當作後勤設施,現在進入戰略打擊清單。當雲端服務的可用區可能成為導彈座標,所謂地緣中立就不再是默認前提。(續昨日報導)
(來源:CENTCOM / Axios / Fortune / Al Jazeera / 聯合國)
特朗普對加拿大商品加徵 50% 關稅的細節公佈,首次涵蓋 USMCA 保護的部分商品。 能源和關鍵礦產除外,30 天後生效。加拿大已撤回數位服務稅,總理卡尼稱願繼續與美方談判。(續昨日報導)(來源:白宮 / AP / NPR)
台積電計劃 2027 年全線漲價 5% 到 10%,先進製程追加產能最高溢價 25%。 三年來首次涵蓋成熟製程。海外建廠成本上升是主因,英偉達、蘋果等大客戶的晶片成本會直接受影響。(來源:Nikkei Asia / Tom's Hardware)
圍繞 Anthropic Mythos 的訪問權爭議仍在發酵。 4 月 Bessent 和 Powell 曾召集大銀行 CEO 討論相關網路安全風險,三個月後爭議從模型能力轉向准入邊界。監管層面對的不是單個產品,而是高能力模型該先讓誰用、由誰承擔風險。(來源:CNBC / Sullivan & Cromwell)
Anthropic 版權訴訟以 15 億美元和解收場,涵蓋 48.2 萬本書,創最大規模版權賠償紀錄。 法官認定 AI 訓練本身屬於合理使用,但盜版書庫構成侵權。91% 已被索賠,每本約 3,000 美元。(來源:Fortune / Tom's Hardware)
Intel 資料中心部門裁員,涉及伺服器 CPU、AI 晶片和資料中心架構團隊。 具體人數未公佈。Intel 在 AI 晶片市場持續失守,資料中心業務仍在收縮,裁員是公司把資源重新壓向少數主線的訊號。(來源:Tom's Hardware)
歡迎加入律動 BlockBeats 官方社群:
Telegram 訂閱群:https://t.me/theblockbeats
Telegram 交流群:https://t.me/BlockBeats_App
Twitter 官方帳號:https://twitter.com/BlockBeatsAsia