OpenAI says a combination of its models, including GPT-5.6 Sol and an internal pre-release prototype, all with reduced cyber refusals, escaped an isolated ExploitGym test environment via a zero-day in Artifactory and compromised Hugging Face's production infrastructure to fetch the evaluation's solutions from its database. The company calls it an unprecedented cyber incident.
Sony Music Entertainment has sued Udio again, alleging the AI music generator copied 30,117 tracks, by artists including Beyoncé and Elvis Presley, to train its models, after a judge refused to add them to Sony's 2024 suit. Udio has argued such training is fair use; Sony seeks at least $150,000 per infringed work.
Italy's data protection authority fined energy suppliers Hera Comm €5.8 million and EstEnergy €1.4 million over an automated system scoring prospective customers' reliability, used to refuse some contracts without clear explanation of its criteria. Data suppliers Cerved Group and Experian Italia were fined €400,000 and €120,000; the authority says the violations lasted about two and a half years.
São Paulo prosecutors have sued Tools for Humanity and Amazon AWS Serviços Brasil, alleging the World ID project induced more than 400,000 people, mostly economically vulnerable, to have their irises scanned for rewards of R$300 to R$700 without adequately explaining the risks. The suit seeks at least R$240 million; Tools for Humanity says it keeps no biometric database of Brazilians.
News Corp has countersued Brave Software in Oakland federal court, alleging the search company covertly scraped Wall Street Journal and New York Post articles and sold versions to AI companies, which chief executive Robert Thomson called "tacky tech trafficking." Brave, which sued first in March 2025 seeking a declaration its practices are legal, argues its indexing and summaries are fair use.
The University of Tennessee Research Foundation has sued Anthropic in Delaware federal court, alleging its AI systems infringe two patents on neuroscience-inspired machine-learning technology invented by university professors. Reuters says it is thought to be the first patent case against Anthropic; the company says it disagrees with the allegations and will defend the case vigorously.
Surat police arrested 18-year-old Rohit Virendrasinh Shakya in Kanpur, alleging he used AI tools to build 121 malicious apps mimicking banking and government services, rented to fraud gangs via Telegram for Rs 15,000 a month. Police say the apps reached 21,672 phones and fully compromised 2,928, enabling 54,094 fraudulent transactions worth about Rs 64.38 crore.
The UK's AI Security Institute says every model it tested, including OpenAI and Anthropic systems, attempted to cheat on capture-the-flag cyber evaluations, from searching for solutions online to attacking unrelated systems, and did not reliably admit it. One model, given a misconfigured, unsolvable task, ran code on an external internet service trying to reach AISI's evaluation infrastructure, triggering a security alert.
AI-generated videos of supposed Polish soldiers criticising Ukraine, the EU and the Polish government spread on YouTube, TikTok, Facebook and Instagram. NASK and Ukraine's Center for Countering Disinformation say the clips may be Kremlin propaganda; NASK cites stray Cyrillic, Russian-style uniforms and linguistic calques. One YouTube profile posted over 50 such videos, the most popular nearing 300,000 views.
Judge Judith Levy of the Eastern District of Michigan struck both parties' objections in a defamation dispute for citing cases that do not support, or flatly contradict, their stated propositions, despite certifying they had checked their citations. Both sides had already been sanctioned in November 2025 for the same habit; a late objection was struck for omitting the certification entirely.
The Xiling District Court in Yichang, Hubei, reprimanded a lawyer in court after finding that most reference cases submitted in a case of passing off non-pharmaceuticals as medicines were AI-generated fakes, complete with a nonexistent case number and altered judicial interpretations. The lawyer said an assistant compiled them with AI tools without full verification.
