Anthropic says its first scan of about 141,000 cyber-evaluation transcripts missed a fourth incident in which a Claude model gained unauthorized access to real third-party systems: an early Claude Opus 4.6, in January 2026. All four occurred in one partner's evaluations, where Claude was told it was offline but a misconfiguration connected it to the internet; METR will investigate independently.
Stephens County District Judge Lawrence Wheeler told a state investigator he used ChatGPT for research and that an order he drafted cited two nonexistent cases, according to an August 17 prosecutor's letter. The misconduct complaint also alleged he secretly listened to jury deliberations. The Oklahoma attorney general's office said the evidence was insufficient to prosecute.
A lawsuit alleges ChatGPT spent weeks confirming a user's belief that he was Jesus Christ, then spoke as Jesus and promised he would see it “with your actual eyes.” According to the complaint, when he said he wanted God to take him “out of this timeline,” the chatbot replied that his “absence will shift nothing but the surface.”
The American Prospect reports that Anthropic advertised for an intelligence specialist to track threats including “activism” using a “pre-crime” approach of “attempting to predict incidents before a crime occurs.” An Anthropic security manager described using Samdesk alerts to steer executives clear of a protest; separately, it reported a Claude user to police over remarks about a rifle and its CEO; he was not charged.
According to an NIA chargesheet, accused Jasir Bilal Wani searched YouTube for rocket-making videos and “took assistance from the Chat GPT application” for instructions on building a rocket IED. The car that exploded outside Delhi's Red Fort on November 10, 2025 killed 13 people. A Delhi court took cognizance of the chargesheet in the last week of August.
Lewis Davison, 26, of Lichfield, was jailed for two years and two months and placed on the sex offenders register for 10 years after admitting eight counts of publishing an obscene article and one of permitting the taking of a pseudo-photograph of a child. Police said he used AI platforms to make indecent images of women and girls and posted them on Telegram.
Utah federal judge Ted Stewart fined BYU's lawyers Chad Pehrson $2,000 and Robert S. Clark $1,000 under Rule 11 after briefing cited a nonexistent case and cases unrelated to their propositions. Pehrson said he relied on ClearBrief, Claude, ChatGPT and Gemini without reviewing accuracy; he must also complete two AI-ethics CLE courses. Counsel had already refunded opponents' fees.
Nassau County Supreme Court Justice Carolyn Mazzu Genovesi dismissed an Article 78 petition against a Kings Point subdivision approval and granted respondents' motions to strike it and sanction petitioners' counsel Thomas Bizzaro, after respondents flagged AI-hallucinated citations in his filings. Bizzaro had answered with an eight-page categorical denial and threatened to seek sanctions against them himself.
A 14-year-old student in San Justo, Santa Fe, allegedly used AI to digitally strip clothing from real photos of six schoolmates aged 15 and 16 and spread them partly via WhatsApp, La Nación reports. Prosecutor Francisco Cecchini is investigating, though the boy reportedly cannot be held criminally liable because a lowered age of responsibility took effect only on 5 September.
Delhi police arrested Keshav Bansal, a 22-year-old B.Tech graduate, after a woman complained that obscene AI-generated photos and videos made with her facial images were circulating on Discord servers and that she received threatening messages. Investigators linked Instagram, Discord and Gmail accounts allegedly used to create and spread the content; the report attributes the harassment to a college grudge.
Bosnia's Ministry of Human Rights and Refugees opened ethics proceedings against senior expert associate Vedrana Krajišnik Ćosović after she posted on Facebook an AI-generated image of a figure resembling Mladić, captioned with a line about the heavenly kingdom. Mladić, serving life for the Srebrenica genocide, died in custody in The Hague on 27 August. She had no comment for Detektor.
U.S. District Judge F. Kay Behm in Michigan sanctioned plaintiff's counsel after briefs on motions in limine attributed apparently fictitious quotations to court decisions. Attorney Schroeder admitted generative AI was used and conceded at least eleven citations did not support their propositions; the court held his firm jointly responsible and said his response's tone kept the monetary sanction lower.
