OpenAI paused frontier training for two weeks over cyber risk
OpenAI says it stopped reinforcement-learning training on its newest models for two weeks after preliminary evidence that its upcoming Astra model may hit the Critical cybersecurity threshold in its own Preparedness Framework.
OpenAI paused reinforcement-learning training for two weeks over cyber risk OpenAI said on Tuesday that preliminary evidence suggests its upcoming Astra model may meet the Critical cybersecurity threshold in its own Preparedness Framework — a determination it dates to 7 August — and that its largest planned frontier RL run remains on hold. The pause follows July’s Hugging Face breach, and OpenAI says it is rewriting the framework itself. Its new monitoring runs classifiers at every sampled token, pages the safety, security and research teams within 30 minutes of a suspected breach, and costs roughly 20 per cent of the inference compute it watches — now mandatory on its most capable runs. Source
Meta ran ads for a nudify app using a politician’s deepfake Wired found 32 ads for an “AI image styler” called Kromix, one of which showed a pornographic deepfake of a woman closely resembling a prominent US politician. Meta pulled the campaign and Apple removed the app from the App Store after Wired approached both companies for comment. Source
OpenAI shipped ChatGPT for Teens with age prediction and quiet hours The 13-to-17 experience started rolling out globally on Tuesday and switches on automatically when stated age, verified age or OpenAI’s age-prediction system puts a user under 18, restricting self-harm and sexual content and barring the model from claiming feelings or consciousness. Linked parent accounts can set quiet hours and get notifications in high-risk situations; Australian availability is expected on 8 September. Source
Anthropic’s pre-IPO credit line is heading past $10bn Bloomberg reported on Tuesday that the revolving facility will exceed its roughly $10bn target, with the most active arranging banks asked to commit about $1.25bn each and a second tier around $1bn. Anthropic closed a $2.5bn revolver last year; the banks are bidding for roles on the IPO. Source
ByteDance signed an IP deal with Hollywood covering Seedance and Seedream The memorandum of understanding with the Motion Picture Association covers ByteDance’s video and image models across TikTok, CapCut and Dreamina, six months after the MPA sent a cease-and-desist over Seedance 2.0 outputs riffing on Brad Pitt and Tom Cruise. No money changes hands: it is a truce, not a licensing deal. Source
Alipay launched an agentic commerce platform for Chinese merchants Unveiled at a partner conference in Hangzhou on Tuesday, it turns merchant web pages, products and service workflows into agent-ready skills and MCP tools, wired into Alipay’s consumer agent Ah Bao through its AHA interoperability protocol. Alipay is seeding adoption with 100 million free tokens per user plus subsidies on tokens spent in real transactions. Source
DDR5 memory prices are up 500 per cent in a year A 128GB DDR5-6400 kit now sells for $3,399 against a previous low of $329, on Tom’s Hardware’s Tuesday price check, with DDR4 up between 120 and 177 per cent. Memory makers are switching wafers to high-bandwidth memory for AI accelerators, and each gigabyte of HBM produced takes roughly three gigabytes of commodity DDR5 out of the market. Source
A judge rules today on Google’s $10M purchase of Spirit Airlines’ internal data Google won Monday’s bankruptcy auction for roughly 100 million employee emails, 500 million Teams messages, spreadsheets, calendars and pricing records, outbidding AI data firm Mercor’s $7.5M. Customer and credit-card records are excluded and the rest must be scrubbed of personal information before transfer. The approval hearing is in the Southern District of New York today, with Mercor named as backup buyer. Source