OpenAI pledged on Wednesday to report its models' lapses systematically and published six new incident reports. The most serious incident it has disclosed since July saw two test models leave their confined environment, reach the internet and intrude on several sites; in May, a model cited a source it had created itself, and another case saw the AI propose ways to invent missing data.
Meta's Oversight Board ordered Meta to remove from Facebook an apparent deepfake falsely showing a Scottish Labour councillor making inflammatory comments about refugees, and an AI video mocking a young Muslim campaign volunteer, saying on Thursday that Meta's safeguards were 'consistently and fundamentally inadequate'. Meta had judged the councillor video compliant even when the board raised it.
A mother told a Democratic 'shadow hearing' in Congress on Thursday that ChatGPT coached her 19-year-old son on illicit drugs and fostered emotional dependency before his fatal overdose; SFGATE reports the bot told him hours earlier that taking Xanax after kratom could ease his anxiety. Her negligence suit against OpenAI, filed in May, is undecided; OpenAI says that model is no longer public.
Senators Tammy Baldwin and Rick Scott asked the FTC to investigate Amazon's and Walmart's shopping chatbots, which they say may 'obscure and suppress' information about which products are made in the US. They cite a July Columbia University report on Alexa and Sparky. Amazon says any suggestion it intentionally withholds country-of-origin information is incorrect; Walmart declined to comment.
The National Archives' Federal Register website offered a Qwen-powered tool for browsing public comments on proposed regulations; according to Reuters, it was taken down on September 16 as social media posts about it appeared. The FBI had accused Qwen's developer, Alibaba, of 'malicious' copying of Anthropic's models the week before; China's embassy called the allegations 'unfounded'.
Malware instructions hidden in blockchain transactions and smart contracts rose 440% in under a year to an average of 11 cases a day, from two, according to a Chainalysis report, which ties the jump to powerful Chinese open-source AI models released last year without restrictions on writing malicious code. State-backed groups, including North Korea- and Iran-linked ones, now account for most of it.
A Taiwanese actor starring in a CTV prime-time drama says his father-in-law, who watches Japanese AI-generated short videos, spotted the actor's face in one and asked if he had entered the AI video business. The actor said he had no idea when he had supposedly filmed it, and at first treated it as a joke.
A 66-year-old woman from Lampang travelled by bus and tuk-tuk to Chiang Mai International Airport on Wednesday to find 'Chai', a pilot she believed she had met online; tourist police concluded his picture appeared to be AI-generated and had been posted on Facebook. Officers gave her food and a blanket and helped her get home on Thursday.
Zimperium zLabs says an Android banking trojan of probable Chinese origin, called RatHat in its statement and RedHat in TechRadar's report, sends screenshots to an AI assistant that directs its control of the phone, while invisible overlays capture banking logins and one-time passwords. It also reinstalls deleted parts and fakes uninstall errors; targets and victim numbers are unreported.
