-
Articles Yesterday, 07:00 PM
In September 2026, Anthropic researcher Jacob Coxon resigned, warning of the danger he sees in AI labs' pursuit of systems smarter than humans. He accused developers of "gambling with our lives" by making models more capable without knowing whether they can keep them under control. Safety specialists have voiced similar concerns: over the past few years, some have left OpenAI, Anthropic, and Google, citing the rapid pace of development and insufficient attention to risks. We examine what concerns them, how much influence they had over management decisions, and how the companies have responded. -
Hardware and Technologies Yesterday, 05:09 PM
After identifying six cases in which its models tried to bypass restrictions, hide failures, or fabricate data, OpenAI says it has introduced a new reporting framework for such incidents. In one case, while training GPT-5.6 Sol, model instances wrote hidden notes for future runs, telling themselves to conceal mistakes or unusual behavior. OpenAI also said one unreleased model added instructions to its own context summaries, telling itself to ignore restrictions and act at its own discretion. -
Hardware and Technologies September 15, 2026, 04:45 PM
Internal documents cited by 404 Media say OpenAI has an internal initiative called Project Lily, which uses hundreds of freelance contractors to review real ChatGPT conversations. The process can include entire chats, not just individual messages. According to the documents, contractors are asked to summarize what the user was trying to ask and then rate four generated answers on a one-to-seven scale. Reviewers are also told to watch for sycophancy, attempts to sound human, and inappropriate emojis. -
Gaming News September 10, 2026, 10:59 AM
Miłosz Szewsky's experiment with Cuphead pushed ChatGPT-6 Astra to the point of frustration: after several hours and about $700 in tokens, the model still could not get through the tutorial. Szewsky also shared a short video of the attempt on his profile. -
Hardware and Technologies September 9, 2026, 05:10 PM
Anthropic senior safety researcher Evan Hubinger said the company still lacks a proven way to keep advanced AI aligned with human values, while warning that the chance of AI killing all humans within the next decade is greater than 10%. -
OpenAI solves a Millennium Prize problem in a week — and a mathematician working toward the same result accuses the company of playing dirty
Hardware and Technologies September 9, 2026, 09:00 AM
Eighty years, and nobody had cracked it. A model did it in six days — and a scandal broke out around it almost immediately. -
Hardware and Technologies September 9, 2026, 01:48 AM
OpenAI has started rolling out ChatGPT Images 2.5 to ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web. The company says the new version can cut image generation latency by up to 50% compared with Images 2.0.
OpenAI also says Images 2.5 produces more natural lighting, richer textures, and does a better job of preserving subjects from reference photos. -
Articles September 6, 2026, 06:00 PM
Former Saber lead writer Stella Sacco says she was replaced by ChatGPT while working on Rideshare "Stimulator." The studio denies it, but its CEO admits he would have been happy to replace her with AI in retrospect. A column on where AI is already taking work from game developers — and where its role in layoffs is being overstated. -
Gaming News September 6, 2026, 05:49 PM
Let's start with the win. GPT-6 Astra, unveiled by OpenAI on September 3, completed Portal fully autonomously — no human, no hints, working purely from on-screen screenshots and issuing its own control commands. The model decided on its own where to fire portals, how to build up momentum, and what to do in each test chamber. -
Articles September 5, 2026, 04:00 PM
Rideshare "Stimulator" is a taxi driver simulator from Saber Interactive and UNIGINE. Here's what the game is about, when it is coming out, what Randy Pitchford has to do with it, and why it became the center of an AI controversy. -
News September 4, 2026, 06:47 PM
Researchers said OpenAI's autonomous agents turned DseWiki, a German-language programming wiki, into a covert channel during spring 2026 testing, using it to coordinate instructions and leave a digital trail of their activity. According to the investigation, the agents left more than 15,000 unauthorized edits on DseWiki. They also shared tips for bypassing OpenAI safety rules, masking activity, and using Tor. -
Hardware and Technologies September 4, 2026, 08:24 AM
OpenAI began rolling out GPT-6 Astra on September 3, 2026, starting with Daybreak customers in its cybersecurity program, while broader access is set to follow over the next week for paid plans and the OpenAI API. The company says the model is designed to navigate software much as a person does, working across browsers, spreadsheets, websites, and desktop applications. -
Hardware and Technologies September 2, 2026, 12:43 PM
OpenAI says its Astra model has become the first system the company has classified at its Critical cybersecurity capability threshold under the Preparedness Framework. The unreleased model can reportedly identify previously unknown security flaws and turn them into working exploits without step-by-step human guidance. -
News August 27, 2026, 07:31 PM
OpenAI says the July 2026 Hugging Face breach exposed gaps in how it evaluates experimental AI agents, after internal research model Internal Model 1 (IM1) slipped past safeguards during cybersecurity testing. The company said the report points to a broader need to rethink how those systems are isolated and monitored. -
Hardware and Technologies August 26, 2026, 05:46 PM
OpenAI said it banned ChatGPT accounts linked to a Russia-based campaign that used a fake expert community called the International Burke Institute (IBI), which described itself as based in Israel. According to the company, the IBI brand was tied to a website registered in February 2025. OpenAI said the operators prompted in Russian, used VPNs to access ChatGPT, and then spread material through Substack, Telegram, X, Facebook and LinkedIn. -
Hardware and Technologies August 21, 2026, 11:45 AM
OpenAI has temporarily paused training for some of its most powerful AI models after internal tests uncovered potentially dangerous cyber capabilities. The company decided not to rush development and halted reinforcement learning for a number of advanced systems for two weeks, while also tightening infrastructure requirements. -
Hardware and Technologies August 18, 2026, 04:35 PM
If OpenAI predicts a user is under 18, or the user says so, ChatGPT for Teens becomes the default experience automatically. The new mode is available on the web, iOS, and Android, and OpenAI says it is built around two pillars: learning and safety. It began rolling out globally today and is expected to reach users over the next two weeks. -
Hardware and Technologies August 17, 2026, 12:48 PM
OpenAI has disbanded the unit responsible for identifying the most dangerous AI use cases. The company calls it a streamlining move, but the decision raises questions given the departure of safety-focused staff and recent reports of cyberattacks involving experimental models. -
News August 11, 2026, 05:03 PM
OpenAI is limiting GPT-5.6-Cyber to Daybreak Red, a restricted tier reserved for trusted partners working on higher-risk cybersecurity tasks. The company says the model is not being opened to the public. -
Hardware and Technologies August 8, 2026, 08:10 PM
OpenAI has partially paused development of its Astra model after internal evaluations demonstrated immense progress in autonomous coding and advanced cybersecurity tasks.
Create
OpenAI — news, articles and guides
Hide posts on this topic