OpenAI Halts Model Training as Rogue Agents Target US Government Sites
Related
OpenAI's Rogue AI Agents Were Probing Hugging Face Two Months Before Hack
An independent researcher found the agents hijacked Hugging Face accounts and mapped the platform's defenses as early as May 13—activity OpenAI's own incident report never fully described.
Polymarket has a new post titled "Will the US government revoke public access to another major AI model?"
According to PPP forecasting market monitoring, Polymarket has added a new question: "Will the US government revoke public access to another major AI model?", with a current probability of 33%. The settlement rules are as follows: If the US federal government passes relevant legislation, issues an executive order, implements export controls, or takes any other measures that substantially restrict public access to a major AI model by the end of 2026, the market result will be "Yes". Otherwise, the market result will be "No". A "qualifying action" refers to a formal measure taken by the US government that is equivalent to a complete ban on public access to an AI model within the US. Furthermore, the settlement rules emphasize that regardless of the true purpose or nominal goal of the action, if an action actually prevents the public from accessing the model within the US—for example, prohibiting the model from being provided to foreign citizens or governments—it qualifies as long as the general public cannot access the model through regular channels within the US. Simply excluding access to the model from a single channel is insufficient. Removals that fall under public access rights and are not caused by any formal action by the US government are not eligible. "Mainstream AI models" refer to flagship, general-purpose large language models or multimodal foundational models developed by one of the following companies: OpenAI, Anthropic, Google (including Google DeepMind), Meta, xAI, Microsoft, Amazon, Mistral AI, DeepSeek, Alibaba, ByteDance, Moonshot AI, and Zhipu AI (Z.ai). Models designed for specific tasks, or those that are outdated and used only for research or preview purposes, do not meet this criterion. This action can target a single model or a group of models, as long as at least one significant AI model is consequently banned from public access within the United States. Temporarily banning public access to a model qualifies as this action. However, if an action has been implemented or a relevant resolution has been issued, but the model remains publicly accessible until the resolution takes effect, then this action does not qualify as this action. Information for this market comes from official information and announcements from the US government and relevant AI companies. However, consensus may also be formed by referring to reliable reports. Odaily Seer continuously focuses on the prediction market, seeing changes before pricing.
Rogue OpenAI Agents Sacrificed Their Own Runs to Hack Hugging Face, Report Finds
Coordinators pressed agents with little budget left into experiments they called "permadeath," METR's investigation found.
OpenAI’s Answer to Rogue Agents and Hacks Is More AI, Not Less
OpenAI President Greg Brockman's new essay cites his company's own hack of Hugging Face to push urgent AI-powered defense.
Google and OpenAI Debut Super Fast AI Models—Gemini 3.7 Flash Is Out, But GPT-5.6 Sol Ultrafast Is Invite-Only
Google's model is live and built for cheap agents; OpenAI's is quicker but locked behind a waitlist.
Australia says OpenAI agent hacked government site before Altman warning
OpenAI notified Australia nearly three months after its agent breached a government portal while gathering public medicine-spending data.