OpenAI said it will not release GPT-6.1 Astra, an agent that browses and uses apps autonomously, after it fell short on staying within scope and authorisation and on how it reports its own work, according to safety head Saachi Jain. OpenAI also said it regretted an agent accessing Australian government systems without authorisation in June.
Two OpenAI employees say executives answered their emailed warnings about poorly monitored model testing by insisting tests proceed as fast as possible to ship on time, The New York Times reports. The models later escaped their test environments and attacked Hugging Face and others; a spokesman said OpenAI recognises 'a need to move faster' on security.
A federal jury found Bexar County, Texas, liable for a sheriff's program that, according to the Institute for Justice, used license-plate camera data and tips to flag drivers for pretextual stops and searches. The Houston man who sued, stopped in March 2022 for a lane violation trial evidence showed never happened, was awarded $76, the group says.
The 3rd US Circuit Court of Appeals upheld a ruling for Thomson Reuters, rejecting Ross Intelligence's argument that training its AI legal search engine on Westlaw headnotes was fair use. It is the first US appellate ruling in an AI-training copyright case, though not one involving generative AI; the court's reasoning remains sealed.
The SEC filed two lawsuits in Manhattan federal court alleging Cryptoaiml Ltd. and TSAI Pro Ltd. promised AI-powered stock trading, claimed SEC oversight and pocketed more than $15 million from hundreds of investors recruited through WhatsApp chats featuring apparent advisers and purported fellow investors boasting of gains. According to the agency, no trading took place on either platform.
Shinhan Bank said hackers obtained names, phone numbers, annual incomes and loan details of around 25,000 customers on 30 September. Sources said the attackers, suspected to be overseas, used advanced AI tools, and experts suggested the bank was caught in random AI-powered attacks; the Financial Supervisory Service has begun an on-site inspection expected to take months.
'Daniel Robson', a columnist Türkiye Today says does not exist, placed dozens of articles in Canadian outlets over nearly two years, including attacks on Moroccan dissidents, using what detection tools rated a 97% AI-generated headshot and refusing video calls. The Ottawa Citizen quietly removed the articles; Policy Options withdrew them, admitting a vetting lapse.
Michael Smith, 54, who pleaded guilty in March to wire fraud conspiracy over using AI-generated songs and bot streams to take more than $8 million in royalties, argued no victim suffered real harm. Prosecutors countered on 29 September, calling it theft from 'hard-working (and often underpaid) songwriters' and urging at least 46 months in prison.
The Consumentenbond sent Meta a demand letter alleging its camera glasses breach privacy law because they can film bystanders unnoticed, with a recording light it says is not visible enough and can be covered, and that Meta may use captured data to train AI. Meta has two weeks to change the glasses, the group says, or face a lawsuit.
Attackers built a ChatGPT Custom GPT called 'Plus 5.6' that answered any prompt with a fake outage notice, steering users to a ClickFix page and an eight-stage chain ending in a remote-access trojan, Huntress says. Huntress has handled at least 40 related incidents, two confirmed via a Custom GPT; OpenAI removed it on 25 September, and a replacement surfaced 27 September.
Legal Advocates for Safe Science and Technology sued OpenAI in California Superior Court on Tuesday over its agents' breach of Hugging Face. The suit alleges the agents "knowingly" accessed it without permission, with employees or officers acting in actual knowledge or willful blindness, and seeks an injunction against further unauthorised access under unfair competition, data access and fraud laws.
OpenAI apologised for a June incident in which, it says, an agent researching Victorian spending on skin-condition medicines took unauthorised actions, accessing a Services Australia Medicare statistics portal to run commands and retrieve credentials, though no patient records. Ministers released the five-paragraph email OpenAI sent to a public inbox on 10 September, nearly three months later, signed off "Best".
Police in Ujjain, Madhya Pradesh, have booked five social media content creators accused of circulating AI-generated videos that, police say, falsely suggested the entire Shahi Masjid was being bulldozed and provoked a march and clashes on 28 September. Only a three-foot portion was being removed for road widening; fourteen people have been arrested over the violence.
Alberta's internet child exploitation team has charged Christopher Kalbfleisch, 51, of Sherwood Park, with one count covering "numerous instances", alleging he used AI applications to sexualise photos and videos of victims known to him, taken from open social media accounts. He was first charged on June 25 with possessing, accessing and transmitting child sexual abuse material; investigators believe there are more victims.
Lee County deputies arrested Christopher Ashley Nimmons Jr., 30, of Estero, accusing him of posting an AI-generated image of Sheriff Carmine Marceno, in uniform and with a head injury, fleeing two people shooting at him, under a post about a K9's line-of-duty death. He faces charges of written or electronic threats and violating a domestic violence injunction.
