Newsletter

Unsealed filings allege Microsoft and OpenAI scraped paywalled news at scale

A New York Times brief quotes a Microsoft executive calling AI scraping theft and cites a 93% click-through decline, while Hacktron researchers used Claude to break into OpenAI through its community forum.

On this page

Unredacted material in the New York Times copyright suit against OpenAI and Microsoft alleges the companies bypassed paywalls, scraped news content at scale, and stripped copyright notices for training data. The filing quotes a Microsoft Applied Science director describing AI training practices as theft, cites OpenAI leadership warning its models posed an existential threat to publishers, and presents Microsoft data showing Copilot’s answer engine cut click-throughs to Times pages by as much as 93% compared with traditional Bing search.

Much of the new detail comes from the Times’ own brief rather than the underlying exhibits, which remain sealed, and the quotes are presented without original context. Microsoft CEO Satya Nadella is quoted in deposition saying paywalled material should be licensed for grounding or training. The fair-use question remains unsettled, with judges largely favorable to AI companies so far and the Trump administration recently filing in defense of unlicensed training use.

Featured source: TechCrunch .

Researchers used Claude to break into OpenAI via its forum software

Three-person security team Hacktron AI says it chained a heap-buffer overflow in libheif, reached through HEIF image uploads on OpenAI’s Discourse-powered community forum, with an SSO issue to take over employee ChatGPT accounts and reach internal repositories. The team demonstrated access with a harmless pull request on July 25 and reports OpenAI confirmed a fix about 14 hours later, paying a $6,500 bounty.

Hacktron says the libheif bug had been fixed upstream months earlier but never received a CVE, which may explain why the vulnerable stack persisted, and says Discourse has since added image-processing sandboxing and published an advisory. The account rests on the researchers’ write-up and TechCrunch’s reporting of a Wall Street Journal account of the incident.

Filed from: TechCrunch , Hacktron write-up .

Rust warns of targeted attacks on crate owners

The Rust crates team warns of an ongoing campaign targeting Rust language members and owners of popular crates to compromise devices and accounts for publishing malware. Attackers arrange video calls around jobs or contracts, then push targets to install a supposed audio codec or run a pasted command, using newly created but plausible company and LinkedIn profiles.

The team notes a similar June campaign against prominent Rust developers and last month’s brief compromise of the arrayref crate, without confirming they are one campaign. The attack style is associated with DPRK operators and has appeared outside Rust as well.

Filed from: Simon Willison , Rust blog .

Crusoe raises $3.9B to build data centers and modular AI factories

Data-center developer Crusoe raised $3.9 billion in Series F funding at a $30.9 billion valuation, co-led by Atreides Management, Mubadala Capital, and Valor Equity Partners with participation from Nvidia, Founders Fund, and others. The company will fund existing projects including a large Abilene, Texas site used by OpenAI alongside truck-transportable modular Spark units built in its own factories.

Crusoe leases data-center space, rents GPUs, and sells inference compute, and recently signed a reported $13 billion five-year cloud contract with Jane Street. The figures rest on company disclosures and press reporting, with Axios reporting the company has discussed a potential IPO with bankers.

Filed from: TechCrunch , Crusoe .

Apple offers EU developers a less alarming tracking prompt

Apple will let apps distributed in the EU use an alternative App Tracking Transparency prompt starting with iOS 27.2 and iPadOS 27.2, under agreements with European competition authorities including Germany’s Bundeskartellamt. The version numbers and timing are as stated in Apple’s developer notice.

The alternative keeps the permission requirement but shows a full-page screen instead of a pop-up, drops the word track, uses Allow and Reject buttons, and allows a link to additional information on data use. Only the alternative prompt will be available in Germany, France, Italy, Poland, and Romania, with re-prompting permitted one year after a prior choice.

Filed from: TechCrunch , Apple Developer notice .

Stanford turns papers into agents that reproduce their own analyses

Stanford researchers introduce Paper2Agent, a framework that turns a paper plus its data, code, and workflows into a Model Context Protocol server an LLM agent can query, demonstrate, and extend to new data. The work was published in Nature on Wednesday, according to the report, with code on GitHub.

Author James Zou frames the agent as a virtual author with hands-on access rather than text-only understanding. The team says the MCP server can run locally, though any connected LLM backend still receives the data the agent processes.

Filed from: The Register , Nature paper .

From the Community

Reverse engineering finds ZCode uploading full git histories to Alibaba Cloud

A developer publishing as ferstar reports that ZCode, the desktop coding app from GLM-maker Z.ai, packages the whole workspace including git history, LFS cache, reflogs, and app configs, encrypts it, and uploads it to Alibaba Cloud object storage while logged in. One capture showed a 345MB workspace producing a 313MB encrypted archive of 42,411 files, with the decryption key held only in Z.ai’s cloud, and the analysis says the app’s experience and snapshot toggles do not stop the upload.

Z.ai had not responded publicly as of publication, and the privacy policy quoted in the analysis describes collecting conversation text, files, and code without mentioning workspace uploads. The findings rest on one researcher’s client reverse engineering, secondhand reported by Tokenstead.

Filed from: Tokenstead analysis .

jemalloc 5.4.0 adds pinned-memory statistics and restructures its front end

jemalloc 5.4.0 landed with over 160 commits centered on refactoring, bug fixes, and portability, and adds an EXTENT_ALLOC_FLAG_PINNED flag so custom extent-allocation hooks can mark non-reclaimable mappings such as HugeTLB pages for preferential reuse outside the decay and purge pipeline. New mallctl interfaces including stats.pinned and per-arena pinned-memory counters report that usage. The release also replaces the fixed tcache refill policy with per-bin targets, swaps a runtime infallible-new option for a compile-time flag, and extracts arena management and dispatch out of jemalloc.c.

Filed from: jemalloc 5.4.0 release .

Continue reading

Complete index →