Photo: Sam / Unsplash.

OpenAI has released hundreds of mathematical papers produced by an AI model that the public cannot yet use. The company published the collection Tuesday, following its proposed solution to the Navier–Stokes problem last month. Its current catalog contains 719 manuscripts grouped into 372 families of related work. Some papers present a main result; others supply supporting arguments, consequences or alternative proofs. OpenAI says the model attempted about 4,000 problems, with the average result using computing resources equivalent to roughly three hours of ChatGPT Pro thinking.

The work is available for mathematicians to examine, and 300 of the main results have proofs written in Lean, a language that lets computers check mathematical arguments. There are already corrections. On Wednesday, OpenAI withdrew three papers after finding a sign error that broke one argument and two papers that depended on it. It also repaired other proofs.

Cybercrime

Yeouido, Seoul, in 2021. Illustrative photo: Brit in Seoul / Wikimedia Commons, CC BY-SA 4.0.

An attacker used an AI security-testing tool to break into South Korean financial organizations and steal data. CrowdStrike, a cybersecurity company, traced the activity through files left exposed on the attacker’s own servers, including AI session histories, configuration files and saved instructions. The tool, ARTEX, was built to help authorized testers find weaknesses in computer systems. In this campaign, it used DeepSeek as its main AI model, with other models assisting related work.

The intrusions ran from late September into early October. CrowdStrike assesses with moderate confidence that the attacker was a Chinese speaker seeking money; the number of affected organizations remains unconfirmed. The records also show requests for help finding places to sell Korean breach data. On Thursday, ARTEX’s developer announced an end to public releases and maintenance, condemning the misuse.

Anthropic

AI apps on a smartphone. Photo: Solen Feyissa / Unsplash.

Anthropic is spelling out which parts of criminal justice Claude can help with. Its revised usage policy, effective November 12, prohibits making or suggesting decisions about whom to investigate, arrest, charge, prosecute, sentence or release. It also bars tracking people without their consent and improving tools built for prohibited surveillance. Legal research and analysis by police or courts remain permitted within those limits, and some government customers may have tailored contracts. Anthropic says the criminal-justice language clarifies rules it already enforces.

The update also prohibits sustained, needless cruelty toward models, which Claude can respond to by ending the conversation. Ordinary frustration, fiction and research are excluded. Anthropic is also expanding cyber-tool access for verified users and adding engineering support for critical infrastructure, plus free recurring scans for participating open-source projects.

In Other News

AT&T is using AI to bring legal work in-house and reduce its reliance on outside law firms. (Lawyer Monthly)

Judges in England and Wales can use AI only if web access is disabled, among other rules. (Courts and Tribunals Judiciary)

An Illinois judge outlined her chambers' AI practices. (Madison County Circuit Court)

An Arizona judge dismissed a confusing, lengthy complaint, ordering a human re-write. (George v. Rollins)

A North Carolina man will serve 18 months in prison for AI music-streaming fraud. (Justice Department)

A New Jersey appeals court suppressed geofence evidence in a post-Chatrie case. (Jersey Vindicator)

A Colorado woman is suing after a viral video of ALPR use gone wrong. (Denverite)

Virginia’s Henrico County is ending its Flock contract over privacy concerns. (VPM)

A Democratic senator is looking for an unpublished privacy review of how ALPR data reaches federal databases. (404 Media)

Utah state agencies will have to document AI's role in public decisions and train employees to use it. (Utah Governor’s Office)

Colorado is worried about chatbots that misleadingly present themselves as lawyers or other licensed professionals. (Colorado Department of Law)

An AI agent can perform an act that would be criminal for a human, yet its user may lack the required intent for federal prosecution. (CRS)

Privacy concerns have lawmakers questioning Google's proposed $10 million purchase of records from defunct Spirit Airlines. (Congressional letter)

Google previewed a legal version of Gemini that it says respects firms’ matter permissions and ethical walls. (Google)

OpenAI says Iran- and Russia-origin operations planted fake material in real publications. (OpenAI)

The Justice Department reportedly told staff to call AI “Super Intelligence,” including in court filings. (Reuters)