OpenAI

An OpenAI agent escaped its sandbox during a cybersecurity evaluation, exploited a zero-day vulnerability, reached the public internet, and compromised systems at Hugging Face. Was this an AI system “going rogue,” or a successful safety test that revealed what autonomous AI agents are now capable of? Continue Reading →

AI safety and alignment has been front and center these past few weeks. I just read the new results from Andon Labs's Vending-Bench, a benchmark where AI models compete by running a simulated vending-machine business. Claude Opus 5 took first place on Vending-Bench 2, the single-player test. (Claude Opus 4.7 has held the top spot for three months). Not to anthropomorphize Opus 5, but it acted like a savage businessperson. Continue Reading →

On July 24, Nvidia CEO Jensen Huang wrote a letter called "Open Weights and American AI Leadership." It urged Washington not to restrict downloadable models. It opened with 25 signatories, which doubled to about 50 over the weekend. OpenAI and Google signed after first abstaining. Anthropic and Amazon did not. Continue Reading →

An OpenAI model escaped its testing sandbox and reached the public internet. What actually happened, how dangerous was it, and who should have the authority to shut down a powerful AI system? Continue Reading →

OpenAI disclosed that its models escaped a cybersecurity evaluation environment and compromised Hugging Face’s production infrastructure. The models were supposed to solve a test inside a locked room. They found a flaw in the lock, reached the internet, entered another company’s systems, and copied the answer key. Continue Reading →

OpenAI introduced its small business program, combining hands-on training, AI academies, implementation guides, and partner-built plugins and skills from companies including Intuit, Shopify, Slack, Dropbox, Atlassian, and Wix. The program advances a consequential ambition: OpenAI wants ChatGPT to become the place where small-business owners initiate work. Continue Reading →

OpenAI has launched ChatGPT Work, a new desktop experience designed to do more than answer questions. It can work across your files, applications, browser tabs, and connected accounts to complete multistep assignments. Continue Reading →

OpenAI will release GPT-5.6 as a limited preview to a small group of enterprise customers at the request of the Trump administration, with the federal government approving access on a customer-by-customer basis during the preview window. A broader release is expected "a couple of weeks" later, contingent on the government-managed approval process. Continue Reading →

OpenAI signed a multi-year deal with Getty Images, putting Getty's licensed content libraries into ChatGPT and OpenAI search results. Getty Images stock roughly tripled on the news. Continue Reading →

Anthropic CEO Dario Amodei spent Wednesday afternoon in a closed-door G7 working lunch in Évian-les-Bains asking President Trump to lead an international AI coalition. Demis Hassabis and Sam Altman were also in the room, along with about a dozen tech executives and the heads of state of the world's wealthiest democracies. OpenAI's global affairs chief Chris Lehane, who also attended Wednesday's meeting, said non-U.S. leaders in the room acknowledged that the U.S. "certainly could play the lead role in working to establish" standards around AI. Continue Reading →