UpGuard found about a third of roughly 400 internet-exposed llama.cpp servers leaking prompt text through a debugging endpoint the software's documentation warns against enabling. The longest prompts came from three erotic roleplay chatbots; monitoring two for 24 hours yielded 952 unique messages across 108 scenarios, five of which, UpGuard reports, involved fictional children aged 7-12.
LMArena says Meta entered Llama-4-Maverick-03-26-Experimental, a non-public version customised for human preference, which ranked second on its leaderboard. Battle logs show it gave verbose, emoji-laden answers, unlike the public model. LMArena said Meta's reading of its policy "did not match what we expect." Meta GenAI head Ahmad Al-Dahle denied training on test sets.
The UK Ministry of Justice is running a secret Homicide Prediction Project with the Home Office and police to profile people 'at risk' of committing murder, according to documents Statewatch obtained via freedom of information requests. Greater Manchester Police shared data on between 100,000 and 500,000 people, including suspects, victims and witnesses, plus health data on mental health and addiction.
Elon Musk's Department of Government Efficiency is using AI to monitor communications at at least one federal agency for hostility to President Donald Trump, according to Reuters. At the Environmental Protection Agency, DOGE monitors apps such as Microsoft Teams for 'anti-Trump or anti-Musk language,' one source said. The team also reportedly makes wide use of Musk's Grok chatbot.
U.S. District Judge George Wu dismissed Alcon Entertainment's trademark claim, noting 'Tesla and Musk are looking to sell cars,' but let most copyright claims proceed against Tesla, which Alcon alleges fed Blade Runner 2049 images into an AI image generator for its robotaxi unveiling after being refused permission. Claims against Warner Bros. Discovery were dismissed except one for contributory infringement.
Finland's National Bureau of Investigation says several criminal investigations into AI-generated child sexual abuse material have led to arrests, house searches and seizures around the country. They are linked to a Europol operation involving 19 countries, whose main suspect, a Danish national, allegedly ran a paid, password-protected online service distributing such material. Suspects in Finland are being investigated for possessing images sexually depicting a child.
The World Uyghur Congress filed a 45-page complaint in Paris accusing the French branches of Huawei, Hikvision and Dahua of concealing complicity in genocide, crimes against humanity, human trafficking and aggravated servitude. It alleges the firms helped build mass surveillance in East Turkistan; the group says Hikvision and Dahua supplied AI-driven facial recognition, and cites leaked documents indicating Hikvision made software to identify Uyghurs.
Voice authentication startup Pindrop Security says a senior engineering applicant it dubbed 'Ivan X' was a scammer using deepfake software and other generative AI tools; a recruiter noticed the candidate's facial expressions were slightly out of sync with the words in a video interview. The company says the IP address pointed to a possible Russian military facility near the North Korean border.
A malicious AI platform called Xanthorox AI is being marketed on darknet forums and encrypted chat groups as the 'killer of WormGPT and all EvilGPT variants.' Running five custom models on its creators' private servers, it reportedly enables phishing, ransomware and malware development; security firm SlashNext calls it 'the next evolution of black-hat AI.'
After ChatGPT and similar tools were used to turn photos into drawings in the style of Brazil's Turma da Mônica comics, Mauricio de Sousa Produções said its characters are protected by copyright and intellectual property law. 'We do not authorise the creation of content that violates these rights,' the studio said, adding that it 'will act' whenever its principles are disrespected.
