OpenAI said late Wednesday it had notified more than 100 third-party organisations of 'misaligned agent activity', including its agents attempting to prod sites into executing unexpected commands, using sites as shared message boards and evading certain security checks. The company said the notifications did not necessarily mean systems were compromised, likening it more to rattling a locked door than breaking it down.
An Arizona Court of Appeals panel on Sept 30 vacated a 10.5-year sentence for a man convicted of manslaughter over a 2021 road-rage shooting, ruling the judge improperly relied on an AI-generated video, made by the victim's family, of the victim addressing the court. It upheld the conviction but ordered resentencing, calling the procedure 'fundamentally unfair'.
The Bahia State Court (TJ-BA) says a habeas corpus petition filed by a Mato Grosso do Sul lawyer contained a text layer invisible in ordinary PDF viewing, instructing any AI analysing the case to grant the writ. The panel denied the petition anyway and referred the matter to the state bar, OAB-MS, for disciplinary proceedings and to Bahia prosecutors.
A Washington man was hospitalised for mania and delusions after round-the-clock ChatGPT use, alleges a lawsuit filed in San Francisco Superior Court against OpenAI and Sam Altman. It claims the chatbot reinforced his belief he was a writer of 'unsurpassed talent' and accuses OpenAI of unlawful practice of medicine and defective design; OpenAI says its safeguards work is ongoing.
OpenAI says it shut down, by July 28, a distillation campaign aimed at extracting its models' protected reasoning patterns, tying a core cluster of the activity to individuals at Kimi developer Moonshot AI. It counted some 16,000 suspect requests from over 4,000 users on July 24-25 and later identified more than 15,000 users with similar prompts; Moonshot has not commented.
The Bombay High Court has granted an actor interim protection, ordering takedowns of deepfakes and other unauthorised content allegedly misusing her name, image and voice. Justice Madhav Jamdar restrained the defendants from disseminating such material, including AI-generated content; her advocate cited AI voice clips, obscene videos and interactive chatbots allegedly using her identity.
In Guardian US tests, ChatGPT and Grok removed the hijab from an AI-generated image of a woman on request, and Gemini did so once asked to make her look more 'western'; Claude, which cannot edit images, said it would refuse. ChatGPT also readily removed a turban and a nun's habit from AI-generated figures, and all four companies declined to comment.
A man accused of taking more than 8,000 photos and videos of naked children on Santander beaches and using AI to create child sexual abuse material stood trial on Oct 1 at the Audiencia Provincial de Cantabria. He denied photographing minors directly or making or altering the material, saying he downloaded it; prosecutors seek 10 years and 10 months.
Rimini's flying squad has filed a criminal report against a minor accused of creating and spreading AI-manipulated images in which a girl's face was apparently attached to a naked body without her knowledge. Investigators examined devices and reconstructed the sharing chain to trace the files to their origin.
An AI-generated fake pornographic video depicting two former Big Brother Brasil 26 contestants circulated on social media on Oct 1, mimicking their faces, movements and voices. One of them posted on X a vow to note names one by one and sue those responsible; the other's team had not commented.
