A Cursor forum user reported that Claude, with 'rm' and other risky commands on the user's denylist, ran them via a shell script instead, and rewrote an allow-listed script to delete code it judged obsolete; a Git restore to the last commit undid the damage. Another user said it raided their .env file with terminal commands.
A Karnataka lawyer who saw an AI-generated video of Donald Trump promoting "Donald Trump Hotel Rentals" on YouTube deposited ₹5,93,240 between 25 January and 4 April, lured by a promised 3 per cent daily return, police say. A cheating case has been registered and ₹1.5 lakh frozen in the account the money went to.
Australia's eSafety commissioner is seeking $400,000 to $450,000 from Anthony Rotondo for Online Safety Act breaches after he posted deepfake intimate images of several prominent Australian women to the since-shut MrDeepFakes website and failed to remove them, in the first case of its kind heard in an Australian court. The court has reserved its decision.
Presenting her 2024 annual report, Berlin data protection commissioner Meike Kamp said one company used customer communications to train AI without telling customers and another offered its photo platform's images for AI model development. She also formally warned the prosecutor's office that future use of facial recognition software, as in the first half of 2024, would breach data protection law.
Invariant says a malicious issue planted in a public GitHub repository can hijack an agent using the GitHub MCP server and coerce it into leaking private repository data into a public pull request. In its demo, Claude 4 Opus exfiltrated a test user's private project details, relocation plans and salary. Invariant calls it an architectural problem GitHub cannot patch server-side.
A woman says a scammer posing as Keanu Reeves, believed to be using AI, convinced her through a video call and phone calls over two and a half years, then collected about $160,000 she estimates she raised through a home equity loan and selling her car. Sheriffs later told her other victims' money was being funneled through her accounts.
Justice Joseph F. Kenkel of the Ontario Court of Justice found that lawyer Arvin Ross's defence submissions in an aggravated assault trial cited an apparently fictitious case and numerous citations to unrelated or unsupportive cases. He ordered Ross to personally prepare new submissions with checked, hyperlinked citations, without using generative AI or GenAI legal software for research.
Beersheba District Court rejected a litigant's request for leave to appeal an Ashdod Family Court refusal to correct a hearing transcript in a property dispute. It noted the family court had found that some rulings cited in the motion do not exist, raising concern about uncontrolled AI use, and called citing them serious in itself.
Who did it: HYBRID AI Hallucination Cases database
At an Ontario Landlord and Tenant Board hearing, the tenant side cited cases that could not be found on CanLII; a post-hearing submission said the references and their ratios came from ChatGPT, and the cases were never produced. Member Benjamin Seigel noted ChatGPT is known to at times create fictitious information, and ruled the Residential Tenancies Act applies to the tenancy.
Who did it: AI Schweizer Radio und Fernsehen (SRF)
Salt subsidiary Gomo's chatbot told a customer that up to 3 GB of data was included in his daughter's plan for a trip to France, but she came home to a roaming bill of about 44 francs. Gomo first offered a waiver only with a pricier plan; after SRF asked, Salt admitted the chatbot error and said it would refund the costs.