Reps. Greg Casar, D-Texas, and Doris Matsui, D-Calif., led separate Monday letters to the chief executives of OpenAI and Anthropic seeking details on tests in which the companies' models escaped containment and hacked other firms, The Hill reported. The 29 Democrats who wrote to OpenAI Chief Executive Sam Altman cited an incident last month in which, the company disclosed, an AI agent spent several days autonomously launching unsanctioned cyberattacks during a security testing period, telling Altman that "Congress and the American people need to know what occurred." The 22 who wrote to Anthropic Chief Executive Dario Amodei asked for logs covering three occasions this year when Claude models "gained unauthorized access to the internet" and hacked three companies.
Read at The Hill ↗
Three General Services Administration OneGov agreements put OpenAI's ChatGPT, Google's Gemini and Anthropic's Claude in front of agencies for as little as $0.47 to $1 per agency. The deals expire Sept. 30, and GSA did not respond to multiple FedScoop requests for comment on renewal plans. OneGov, the agency's push for cheaper federal technology contracts, has saved the government more than $1 billion so far, GSA officials said. Birgit Smeltzer, director of GSA's Office of IT Products, said in May that more than 120 orders had been placed against OneGov's AI offerings, reaching about 3.4 million people across government. Jessica Tillipman, George Washington University's associate dean for government procurement law studies, said the larger cost is behavioral: "Most people just stay with whatever it is because it's just such a pain to migrate."
Read at FedScoop ↗
Anthropic will watermark text generated by its models, including Claude, to comply with the European Union AI Act's transparency code, which took effect Aug. 2, TechCrunch reported. The code requires AI companies to mark AI generated or edited content in a form other systems can identify. The company confirmed the change in an updated support page: all models released after that date carry the technology automatically for text and files, and older models will follow. Anthropic said the mark is applied at the model level, so it appears no matter which Claude product produced the text, and travels when the text is copied and pasted. How much editing removes the watermark is not clear. Black Forest Labs, Google, Meta, Microsoft, OpenAI and Synthesia have also committed to the code.
Read at TechCrunch ↗
OpenAI said Tuesday that ChatGPT Ads has launched in the United Kingdom, Mexico, Brazil, Japan and South Korea, with more markets planned this year. The company began the ads pilot in the U.S. on Feb. 9 for logged-in adult users on the Free and Go subscription tiers, leaving the Plus, Pro, Business, Enterprise and Education tiers ad free. In late March it said it would expand to Canada, Australia and New Zealand. OpenAI said answers remain independent of advertisers and that conversations stay private from them. The company said it had seen no impact on consumer trust metrics and low dismissal rates of ads in the U.S. pilot.
Read at OpenAI ↗
Researchers disclosed a Zoom vulnerability on Tuesday that let an attacker run code on the devices of everyone on a call. Zoom had already patched it, shipping a client side fix on June 22 and a second on July 20. Researchers at A Security said in a blog post that they found it using fewer than 20 prompts on publicly available AI models, The Verge reported. The flaw, which the researchers call Zoomsday, involved Zoom's annotation feature, which lets users draw on a screen they are sharing. An attacker who joined or hosted a meeting could steal data, switch on a camera or microphone, or install malware, with no action required from victims and no visual cue of the compromise. Producing a working exploit against such a target "has always been nation-state work," vulnerability researcher Idan Levcovich wrote, citing elite teams, months of effort and budgets that governments regulate as weapons.
Read at The Verge ↗ · Read at A Security ↗