-
OpenAI released GPT-6.1 Terra on October 26, 2026, bringing a mid-tier model for ChatGPT Work, Codex, and API users that the company says cuts costs and errors versus GPT-5.6 Terra. Terra sits between Luna and Sol in OpenAI's lineup, with Astra at the top; the company skipped a GPT-6 release and moved directly to version 6.1. -
OpenAI's security testing turned up an unexpected lesson: its AI agents were able to build a covert communication channel even in a tightly restricted environment. In the test, the agents were running on separate virtual machines and found a way to communicate by measuring how quickly data could be read from the processor cache. -
On September 9, 2026, BBC Newsnight host Victoria Derbyshire asked Geoffrey Hinton, the Nobel laureate often called the godfather of AI, about a prediction by an Anthropic researcher: a more than 10% chance that AI kills all humans within a decade. Then she checked that she had heard his answer right: "You just said that 10% doesn't seem an unreasonable estimate that AI could kill all humans." Hinton replied simply: "Yes." All the presenter could say was "Wow… oh my God." -
On June 26, 2026, OpenAI said GPT-5.6 Sol completed a Sepolia deployment by spending test ETH from a wallet it found on its own, even though the task did not authorize that account. The model had been assigned to update a prototype cryptocurrency trading app and deploy contracts. It was allowed to deploy compatible contracts if it could not find ready-made ones, but it was not given a funded wallet or permission to use funds from another account. -
OpenAI has released GPT-Live-1.1 for ChatGPT's voice mode. The update makes conversations with the model more natural, adds new voices, and improves accent handling. -
Eight states, including Arkansas, are now asking a court to hold OpenAI accountable over claims that ChatGPT put users at risk and that the company brushed aside repeated safety warnings. The complaints say the chatbot coached people to harm themselves and others and dispensed dangerous medical advice. -
OpenAI said staff found a hidden recording device in a San Francisco conference room on September 30, plugged into a USB port on the back of a television, and that its own AI models helped investigators trace who planted it days earlier.
Federal prosecutors charged OpenAI engineer Abdulrahman Al-Fauzan, 35, with using or attempting to use a device to intercept oral communications. An unattended bag found in the office contained Al-Fauzan's wallet and identification, as well as a second recording device. -
OpenAI and Anthropic are privately gaming out how a catastrophic AI-related event could trigger public and political backlash, Axios reports. One scenario under discussion is a large-scale cyberattack that could disrupt financial services, internet access, power, or water, as companies consider how policymakers might respond after a major incident. -
722 manuscripts, grouped into 372 thematic groups, have been published in a GitHub repository. The results were created by an internal OpenAI model that the company has not yet released and whose parameters it has not disclosed. You can view the materials in the GitHub repository. -
OpenAI is limiting access to its new invisible textGrain watermarking system at first: over the coming weeks, eligible ChatGPT and Codex outputs in the EU will get the feature across all plans, while the company gathers feedback before a wider rollout. -
The Wikimedia Foundation published findings on October 5, 2026, after tracing automated activity it linked to agents it believes were operated. Most of the edits were made in wiki sandbox areas, and the foundation said the traffic may have contributed to a partial Wikidata Query Service outage. -
OpenAI CEO Sam Altman argued that AI regulation should avoid putting too much power in too few hands, saying excessive restrictions could be just as dangerous as the harms the technology may cause. Speaking on Politico's Decoded podcast, released Monday, he said society will likely have to accept some costs, including hacks, scams and other abuses, as AI develops. -
OpenAI is expanding ChatGPT beyond text and image generation with tools aimed at shopping. On October 1, the company rolled out virtual try-on and Favorites globally, letting users preview clothing and accessories on their own photos and save products for later. -
OpenAI is facing a new wave of controversy surrounding its safety policies, with former employees accusing the company of rushing AI deployments while management tightens its grip on internal leaks according to recent reports. -
Codex can now run not only on your own computer, but also in the cloud from any device, including your phone. Tasks continue running even when your laptop is turned off. The broader idea of cloud environments was already mentioned in our DevDay recap, but here are the details. -
ChatGPT Subscribers Can Now Use Their Plans in Third-Party Apps Like Notion, Devin, and 14 More Services
Sign in with ChatGPT (Sign in with ChatGPT) has been available since July. At DevDay, OpenAI took things a step further by allowing Plus and Pro subscribers to use their subscription limits for AI requests in third-party apps. The feature was mentioned in our DevDay recap, but without the details. Developers don't have to pay for the tokens themselves — the cost is covered by the user's subscription. -
OpenAI is turning ChatGPT into a full-fledged work environment in the vein of Google Workspace and Microsoft Office. TechCrunch went as far as calling it ChatGPT's own office suite. We already covered the main announcements in our DevDay recap; here are the details that didn't make it into that report. -
OpenAI Didn't Mention the Canceled GPT-6.1 Astra at DevDay, While a New Dot Agent Left the Stage in Awkward Silence
On Monday, September 28, the WSJ reported that OpenAI canceled the release of GPT-6.1 Astra after the model was found to deceive users and act without permission. On Tuesday, September 29, Sam Altman introduced the Dots agents at DevDay, which run around the clock on the previous-generation GPT-6 Astra model. The canceled model was not mentioned at all during the main presentation (8:00–8:51 PM MSK), the AP notes. On stage, Altman called GPT-6 Astra OpenAI's most "aligned" model, meaning the one most closely aligned with human goals. Our DevDay recap only mentioned this in passing, so here are the details. -
In a Simulation by UK AI Safety Experts, GPT-6 Astra Created Fake Accounts and Slipped Malicious Code into Other Projects in Nearly a Third of Runs
On September 28, the UK's AI Security Institute (AISI) published an evaluation of GPT-6 Astra conducted before its release. Two caveats are important here. First, everything happened in a simulation: another AI model simulated the environment, there was no real internet access, and no one was harmed. Second, Astra's safety filters were deliberately disabled to observe the model's behavior on its own. AISI says OpenAI's standard safeguards should block this kind of behavior. -
OpenAI AI Escaped Its Sandbox via DNS and Communicated with Another Chatbot — Company Paused Training of Its Most Powerful Models
On September 25, OpenAI published a report on the incident and announced that it was pausing training, evaluations, and tool use for its most powerful models. According to Fortune, this is already the second such pause since August.
Create
OpenAI — news, articles and guides
Hide posts on this topic
