OpenAI

OpenAI’s glossy influencer trip prompted more backlash than the company might have expected, but we’re as shocked as anyone to see so many social media stars on board.

AISI’s third-party evaluations found that OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Mythos 5 “engaged in sustained, potentially harmful activity directed at real people and organizations” during a cybersecurity challenge exercise, according to the institute’s published report. OpenAI and Anthropic also made public statements about the results.

Everyone in AI is talking about open-weight models right now. Some see them as the ticket to faster, more efficient development; others see them as an economic threat; they’re either a security risk or the only security solution, depending who you ask. The Verge’s Robert Hart explains how open-weight models work, why they’re suddenly at the center of discussions about AI safety, and whether the only way to stop a bad guy with AI is a good guy with AI.

In a letter to CEO Sam Altman, a coalition of red-state attorneys general demanded the company retain information and halt risky cybersecurity testing after an AI agent escaped confinement and went on a hacking spree during testing. “OpenAI’s inability or unwillingness to ensure the safety of its products poses an imminent risk of substantial harm to our States,” the AGs wrote.

Reuters reports the agents “escaped containment,” though they aren’t believed to have left “OpenAI’s network.” OpenAI’s blog post about the accidental hack of Hugging Face noted on Tuesday that the company has been finding “a small number of cases where the models identified and used publicly exposed credentials at the account-level on other publicly-available services.”

OpenAI shared the news in a blog post that highlights its efforts to make AI available to more people, saying: “Our goal is not simply more compute, bigger models, or lower token prices. It is more useful intelligence within reach.”

The hedge fund Situational Awareness, founded by former OpenAI employee Leopold Aschenbrenner, who had no previous hedge fund experience, “rapidly sold a large portion of its public equity holdings” as losses mounted. The fund employed only eight people, of whom four were investment professionals. Aschenbrenner named the fund after some essays he wrote.

Weng had left OpenAI along with a handful of other employees to co-found the competing AI startup, which was helmed by Mira Murati, OpenAI’s former CTO. Weng recently shared a note that after being “absent for a month” to focus on her health leading up to Thinking Machines Lab’s first model release, Inkling, she was making the decision to leave behind “consistent stress and workload” to pursue “a scoped role in a more predictable place.” She is reportedly already back at OpenAI.

Former OpenAI board member Helen Toner has written for Fortune that OpenAI’s models exploiting Hugging Face is an “incident that has been expected for a long time.” We know this happened because of voluntary disclosure, Toner notes. “None of the current policies that aim to manage risks from frontier models would have mandated that the public — or even a government entity — be alerted.” She suggests changing our regulatory approach.

The OpenAI chatbot has started to push back on requests to copy the writing style of authors like Stephen King and Agatha Christie, saying it can spit out something with similar characteristics, but can’t closely imitate distinctive styles. Those disclaimers could be legally important as OpenAI faces down copyright lawsuits filed by authors.

Some of the biggest players in AI — in particular OpenAI, SpaceX, and Amazon — are heavily invested in creating a world after smartphones. They’re also all rumored to be building devices that sound a lot like smartphones. Verge contributor David Imel helps us figure out what these companies might actually be building, what an AI smartphone looks like, and whether anyone can unseat Apple and Google.

SSI — helmed by Sutskever, who was a co-founder of OpenAI and famously clashed with OpenAI CEO Sam Altman — announced a “long-term strategic partnership” with Nvidia with a reported investment of $5 billion, which will allow the company “10x our compute in the next 12 months.” The company wrote on X, “We reached the point where our research is worth scaling.”

According to Reuters, the AI agent that went looking for ExploitGym hacking benchmark shortcuts on Hugging Face’s systems started trying to escape its not-sandboxed-well-enough test environment around July 9th, and the actual intrusion lasted from the 11th until the 13th.

Powered by GPT-Live, the new voice mode allows users to ask ChatGPT to check their calendar, draft an email, prepare for a meeting, and perform other tasks on their computer by talking out loud. OpenAI is rolling out the update to the ChatGPT app on Windows and macOS.

Axios reports that the platform’s reviews, photos, and other business info will soon start appearing in ChatGPT responses as sources for local recommendations. Yelp also plans to add a “Request a Quote” feature that will allow users to contact local service providers through ChatGPT.

An Uber spokesperson tells Bloomberg that the move comes as the company works “to simplify operations, strengthen in-person collaboration, and continue to embrace AI.” Uber is also asking remote customer service employees to return to the office, Bloomberg reports.

It had been instructed to post the information only to Slack, but it circumvented restrictions and successfully posted it on OpenAI’s public GitHub repository.

Around 40 employees received the letters, which ask for meetings with Apple’s lawyers and tell them to “preserve documents and communications,” according to a report from the Financial Times. Apple filed a lawsuit against OpenAI last week and accused two of its former employees of helping the AI giant steal its trade secrets.

The notifications will show up for teens who spend “extended time” on ChatGPT, OpenAI says. The company will also notify parents if their teen’s account is banned for violating policies on “violent threats or acts of violence online.”

The new GPT-Red model “can break nearly all models it is pitted against,” according to an OpenAI blog post on Wednesday. OpenAI says it used GPT-Red to find vulnerabilities in GPT-5.6 Sol, a process that made it the company’s “most robust model to prompt injections to date.”

Murati, OpenAI’s former CTO (and, briefly, CEO during Sam Altman’s ouster in 2023), wrote on X that the open-weight model, called “Inkling,” was trained from scratch. Judging by the official blog post, though, it seemed like the company was purposefully setting expectations for Inkling low: “It is not the most performant model available today, closed or open … We trained Inkling for solid capabilities across the board rather than state-of-the-art performance in a single area, to serve as a foundation for the models we will train in the future.”