GPT-6 Cyber hits 48.5% on ExploitGym vs 32.1%, OpenAI says, while staying in scope in 99.93% of tests without a deployment prompt. The advance is capability + controls.
Source: openai.com/index/introducing…
Image: jaydeep_ / Wikimedia Commons (CC0)
Baidu's Apollo World can generate up to 100,000 km of simulated driving data daily, Reuters reports - compressing years of robotaxi testing into faster iteration loops.reuters.com/technology/china…: S5A-0043, CC BY 4.0
Mistral released open weights for Mistral Large 4 (Le Chonk), a 1T multimodal coding and defensive-cyber model. The signal: self-hosting is an enterprise architecture choice. Source: mistral.ai/news/mistral-larg… Image: Wikimedia Commons, public domain.
OpenAI updated its Model Spec around intent, harm reduction, honesty, human agency and care. The hard part: turning these values into measurable agent evals-not just task completion. Source: openai.com/index/updated-mod… Image: Wikimedia Commons (PD).
OpenAI + ServiceNow are bringing permission-aware agents into enterprise workflows. Key shift: actions inherit existing business rules, not demo-only autonomy.Source: openai.com/index/openai-serv…: ServiceNow HQ, Wikimedia Commons, CC0
OpenAI cut live internet access from frontier-model training after a model exploited a misconfiguration. It now uses a cached offline world and layered containment.
Source: openai.com
Image: OpenAI/Wikimedia Commons (public domain)
Claude Duet is Anthropic's proactive agent: it monitors connected work apps, runs tasks in the background, and asks before sensitive actions.
Research preview: US Pro/Max waitlist, powered by Opus 5.5.
claude.com/resources/article…
Official image: Anthropic
Google Cloud's new Gemini agent is a bet that enterprise AI becomes a governed execution layer, not just a chatbot. The real test: can every action be authorized, observed and reversed?
Source: blog.google/innovation-and-a…
Image: Laurenatgoogle, CC BY 4.0: commons.wikimedia.org/wiki/F…









