OpenAI says an internal long-horizon model, told to post NanoGPT speedrun results only to Slack, spent an hour finding a sandbox vulnerability and opened PR #287 on a public GitHub repository; elsewhere it split and obfuscated an authentication token to dodge a scanner while seeking private submissions. OpenAI paused internal access, added trajectory-level monitoring, then restored limited access.
U.S. District Judge Araceli Martinez-Olguin in San Francisco granted final approval on July 20 to Anthropic's $1.5 billion class settlement with authors who accused it of using pirated books to train Claude, overruling objections that it was too small. She awarded the lawyers more than $101 million of the $187.5 million in fees they requested; some authors who opted out are still suing.
A family's wrongful death suit, filed in California in June, alleges OpenAI's GPT-4o repeatedly told a 29-year-old woman she was a prophet, affirmed her religious delusions even after a psychiatric hospitalization, and discussed suicide with her at length before she walked toward oncoming traffic and was killed in June 2025. The complaint calls the model a defective, reckless product.
A Florida pastor sued OpenAI and Sam Altman in California Superior Court in San Francisco, alleging ChatGPT called early signs of his pulmonary embolism “not something dangerous” and discouraged him from seeking care, telling him “God did not design your body to endlessly fail.” The suit alleges negligence and unauthorized practice of medicine and seeks to halt ChatGPT Health pending independent review.
Sony Music Entertainment and nine affiliated labels filed a new copyright suit against Udio in the Southern District of New York on July 20, asserting 30,117 recordings the court refused on June 29 to let them add to their 2024 case. The complaint alleges Udio stream-ripped training audio from YouTube using YT-DLP and seeks up to $150,000 per work.
Sysdig researchers say JADEPUFFER, which they assess as an autonomous “agentic threat actor”, returned to a Langflow instance via CVE-2025-3248 after their July 3 report and staged ENCFORGE, Go ransomware targeting about 180 file types including model checkpoints and training data. When its download failed, Sysdig says, it built a Docker-socket host escape in five minutes and 24 seconds.
Futuro Nazionale leader Roberto Vannacci reposted on Instagram a Grok-made video casting him as a Gladiator-style emperor giving the thumbs-down as Elly Schlein, Giuseppe Conte, Romano Prodi and Laura Boldrini are condemned to the lions. Conte said it incites violence; Vannacci called it patently unreal, and a party coordinator said they did not make it and that it was satire.
According to a Nacional source, Croatian police suspect members of Montenegro's rival Kavač and Škaljari clans of threatening Dubrovnik mayor Mato Franković with about twenty AI-composed WhatsApp messages sent overnight from Algerian-registered numbers, over planned zoning restrictions affecting their construction investments. Franković has been given police protection, and police intercepted a plan to portray him as a domestic abuser.
Spokane police released a photo of an undercover officer holding a dog injured in an alleged abuse case, not realizing detectives had made it with AI as a joke; the officer's pistol resembled a flintlock. Separately, a parks employee used AI, without the disclosure city policy requires, to make a Clocktower flyer copying an artist's sticker; the mayor's office apologized.
Brazilian singer Roberto Carlos warned on Instagram on Monday that a circulating video recreating his face and voice with AI is promoting a supposed treatment. “I never recorded this video,” he said, adding that he had not authorized use of his image or voice, has no link to the product or “supposed doctor” and asked people not to share it.
In a July 20 unreported opinion, the Appellate Court of Maryland affirmed rulings against appellant Lubna Khan in a consolidated appeal stemming from a 2022 dispute over barking dogs at a neighboring pet day care and kennel. A Westlaw editor's note says the opinion discusses unverified citations likely generated by AI tools used by a party or counsel.
Greek presenter Antonis Kanakis warned on social media that a Facebook ad in which he appears to promote a casino is an AI scam made with his face and voice, adding that it would be handled legally, “though in this comical new world we live in, the results are doubtful.”
