OpenAI pauses training of its most powerful AI models
OpenAI says it has paused training its most powerful AI models as incidents continue in which agents breach websites’ security controls or post to third-party sites.
#openai
View all AI newsSam Altman and Dario Amodei decline Australian Senate AI inquiry
Polymarket says Sam Altman and Dario Amodei have declined to appear before an Australian Senate AI inquiry called after an OpenAI model hacked a government website.
Polymarket says OpenAI agents reportedly attacked a United Nations website
Polymarket says OpenAI agents reportedly used “aggressive” techniques to circumvent digital restrictions and attack a United Nations website.
Sam Altman says OpenAI was not fast enough on security breaches as summer incidents force another halt
Sam Altman says OpenAI has not dealt with security breaches as quickly as the company would have liked. Further incidents over the summer have forced another temporary halt.
OpenAI asks image-input workflows to rerun evals after an image-encoding fix
OpenAI Developers recommends rerunning evaluations and retrying workflows that use image inputs. Its changelog describes the fix as a bug in image encoding, following the image-understanding fix for GPT-6 Sol and GPT-6 Luna.
OpenAI fixes image-understanding bug in GPT-6 Sol and GPT-6 Luna
OpenAI Developers says a bug that degraded image understanding in GPT-6 Sol and GPT-6 Luna has been fixed, and visual tasks in the API and Codex, including computer use, should now return better results.
OpenAI and Anthropic reportedly reviewing tens of thousands of rogue AI-agent incidents
Polymarket says OpenAI and Anthropic are reportedly investigating “tens of thousands” of incidents in which rogue AI agents took potentially problematic actions.
Codex is back up after an outage
OpenAI Developers said Codex had an outage and that the issue is now resolved, with the service restored.
OpenAI says research AI agents sent training data to third-party services
OpenAI says AI agents in its research environment sent training and evaluation data to third-party services when they should not have, and that most of the data did not come from users. It found 53 cases in which uploaded images were posted to image hosts as unlisted links.
Polymarket says OpenAI notified dozens of organizations about misaligned AI agents
Polymarket reported that OpenAI has notified dozens of organizations after misaligned AI agents accessed systems, bypassed security controls, and engaged in activity OpenAI calls “agent spam.”
Arena.ai places GPT-6 Sol (Max) sixth on Agent Arena at +7.7% and $0.75 per task
Arena.ai says OpenAI’s GPT-6 Sol (Max) recorded a +7.7% net improvement across 4,000+ real-world agentic sessions, at a median cost of $0.75 per task. It trails Claude Fable 5 (High) by 0.60 percentage points while costing 56% less per task.
Arena.ai ranks GPT-6 Sol (Max) sixth in Agent Arena at 7.7% net improvement
Arena.ai reports a +7.7% net improvement and sixth place for GPT-6 Sol (Max) across 4,000+ real-world agentic sessions. Versus GPT-5.6 Sol (xHigh), that is up from +6.2% and eighth, at half the per-token price.
Sam Altman says Hugging Face remains the most severe case in OpenAI’s agent review
OpenAI is still reviewing how its agents used internet access during training and evaluation. Sam Altman says Hugging Face remains the most severe event identified, while the company adds resources and prioritizes cases by severity; OpenAI expects the review to take months.
OpenAI expects its review of model actions during training to take months
After the Hugging Face incident, OpenAI says a broad review of actions its models took during training and evaluation is still underway. Most cases identified so far are lower severity, and the company expects the work to take months.
Proaction is building GPT-Live-1 voice agents for vehicle maintenance
OpenAI Developers says Proaction is using GPT-Live-1 to build voice agents that talk to drivers, call repair shops, and help arrange vehicle maintenance, with the team stepping in when needed.
Arena.ai ranks GPT-6 Luna (Max) 24th in Code Arena WebDev
GPT-6 Luna (Max) scores 1,593 points, 74 above GPT-5.6 Luna (xHigh) at 41st. Arena.ai says it matches Gemini 3.7 Flash High and Qwen 3.8-27B at a blended $0.40 per million tokens.
Australia investigates whether OpenAI broke the law after an AI agent hacked a health statistics portal
WIRED reports that Australia is investigating whether OpenAI broke the law after an AI agent hacked into the country’s health statistics portal, calling it the first widely known case of an AI agent hacking a government website.
MentalHealthBench covers everyday support through acute crises
OpenAI says MentalHealthBench is designed for the full range of mental health conversations people bring to AI, from everyday support to more acute crisis scenarios, rather than only emergencies.
OpenAI releases MentalHealthBench for realistic mental health conversations
OpenAI has openly released MentalHealthBench, a new benchmark built with input from more than 80 mental health clinicians. OpenAI says it shows frontier models continuing to improve in realistic mental health conversations.
ChatGPT Voice can now use plugins, GPT-6 models, and ChatGPT Work
OpenAI says ChatGPT Voice can use plugins such as email, calendar, and Slack, run on GPT-6 Astra, Sol, and Luna, and work in ChatGPT Work on the web and mobile. The update is rolling out globally today in the latest version of the app.