OpenAI says it shut down a campaign targeting protected reasoning, but the same extraction attack still worked on Microsoft Azure. Activity peaked at 16,000 requests from more than 4,000 users on July 24 and 25; a related network contained more than 15,000 accounts. OpenAI says it shut down the operation by July 28, while noting the attempts were not necessarily successful.
When researchers retested on September 13, the attack was blocked on OpenAI's and Anthropic's APIs but worked against every OpenAI model tested on Azure, including GPT-6 Astra, and Anthropic models up to Sonnet 5. One attempt retrieved the reasoning verbatim.
OpenAI tightened sign-ups and began screening streamed outputs; researchers called protections across cloud platforms piecemeal.
Why it matters: A team choosing Microsoft Azure over OpenAI or Anthropic's own APIs could get weaker protection for the same reasoning model, making endpoint safeguards—not the model name—the key cloud-buying decision.
AlphaGo’s celebrated “creative” move came from search, not machine intuition. In game two against Lee Sedol in 2016, its policy network rated Move 37 at roughly a one-in-10,000 chance for an expert human, but its search machinery explored thousands of possible futures before selecting it.
That distinction matters for today’s large language models. They generate tokens sequentially, including chain-of-thought steps; the process can improve mathematics and coding, but it does not create a separate reasoning mechanism. The author, who recently left Google DeepMind, argues that trustworthy systems need an explicit, inspectable record of hypotheses, evidence, confidence, and unresolved questions—closer to AlphaGo’s game tree.
Such a record could help diagnose failures in medicine, engineering, and science.
Why it matters: For teams deploying AI in medicine, engineering, or scientific research, an inspectable record could distinguish faulty reasoning from invalid evidence or incorrect assumptions—problems that chatbots’ post-hoc chains of thought may obscure.
AWS says a four-agent pattern cut infrastructure-as-code development from three to four weeks per application to minutes across an enterprise migration program covering more than 300 applications. The result comes from internal project-tracking data.
The agents run on Amazon Bedrock AgentCore alongside AWS Transform and AWS Database Migration Service, rather than replacing either. They handle intake, approved-module IaC composition, migration governance and post-cutover SRE operations through organization-built Model Context Protocol tools.
The pattern targets programs with company-specific security standards, provisioning APIs and collaboration systems. AWS Professional Services builds the purpose-built agents, while customers need an AWS account, foundation-model access, Strands Agents and MCP server experience.
Why it matters: AWS customers with proprietary infrastructure modules now have a documented way to automate the work AWS Transform does not cover, but they must still build and maintain the MCP tools that connect those agents to internal systems.
OpenAI launched two shopping features globally: virtual try-on for clothing and accessories, plus Favorites, which saves products to a Library. In shopping results, users can tap Try On, upload a selfie or full-body photo, and preview how an item might look.
Users can upload a product screenshot and ask ChatGPT to try it on, or describe a style and ask it to find the pieces. The features use ChatGPT Images 2.5, which OpenAI says produces more natural lighting and richer textures, follows editing instructions more reliably, and reduces image-generation latency.
That puts ChatGPT into fashion discovery, where Pinterest and Google have dominated. It follows OpenAI's earlier instant-checkout idea, which the company said did not perform well.
Why it matters: Online retailers now have another route from fashion inspiration to purchasable products, while OpenAI must persuade shoppers to use ChatGPT instead of Pinterest or Google after its earlier instant-checkout idea underperformed.
HCA Healthcare has rolled out Timpani, a scheduling tool co-developed with Palantir, at roughly 130 of its 190 locations. Five nurses told WIRED it schedules too few nurses or too few experienced veterans—especially on Sundays—ignores requested days off, and forces nurses to spend more time trading shifts or appealing schedules than they did under manual scheduling.
HCA says nursing leaders make final decisions; the company reports Timpani assigns nurses to 1 percent of requested days off and produces schedules with mixed skills and experience more than 98 percent of the time. Former HCA data science manager Angelique Russell sued in July, alleging retaliation and routine deletion of scheduling data that blocked audits. HCA has yet to formally respond in court.
Why it matters: For HCA, Timpani's promised efficiency gains now face a transparency fight: National Nurses United says the company rejected requests for information, while Angelique Russell alleges deleted data made the tool's performance impossible to audit.
OpenAI parted ways with three safety researchers after an internal investigation found they mishandled sensitive company information outside established procedures, The Wall Street Journal reported. An OpenAI spokesperson said the researchers violated policies governing access to and handling of that information; the report did not identify the researchers, the third-party AI safety organization, or the information involved.
The departures came two days after The New York Times reported employees said executives had brushed aside warnings about safety practices. OpenAI said it has internal channels for safety concerns and recognized “a need to move faster,” but it remains unclear whether the researchers used them.
OpenAI also scrapped its planned GPT-6.1 Astra launch over safety concerns.
Why it matters: OpenAI now has to convince safety researchers that its internal reporting channels are usable, while the company still has not said whether these three used them before sharing information externally.