UN Security Council Convenes Historic AI and International Security Session
In what marks a watershed moment for AI governance, the UN Security Council held a high-level briefing on artificial intelligence and international security today, September 23. Convened by France — which holds the Council presidency this month — the session brought together an unprecedented lineup: OpenAI CEO Sam Altman, Anthropic CEO Dario Amodei, and Hugging Face CEO Clément Delangue, alongside representatives from China's DeepSeek and Moonshot.
This is the first time the Security Council has directly hosted frontier AI developers from both the US and China together to discuss shared safety concerns. French Foreign Minister Jean-Noël Barrot chaired the meeting, with Yoshua Bengio, co-chair of the UN's Independent International Scientific Panel on AI, also briefing the Council.
The session comes at a critical juncture: just last week, OpenAI, Anthropic, and Google DeepMind confirmed they have been coordinating on AI safety standards for several weeks, following Dario Amodei's September 12 essay calling for frontier developers to adopt common safety standards and grant outside evaluators deep access to their systems.
OpenAI Launches GPT-6 Sol and Luna: Better Performance, Half the Price
OpenAI rolled out GPT-6 Sol and GPT-6 Luna on Tuesday, rounding out the GPT-6 family that launched earlier with the flagship Astra model. The new models represent a significant step in making frontier AI more accessible and affordable.
GPT-6 Sol is designed for everyday professional work and costs $2 per million input tokens / $10 per million output tokens — exactly half the price of GPT-5.6 Sol. OpenAI claims Sol makes "about half as many mistakes" as its predecessor, with improvements across factuality, coding, computer use, and alignment.
GPT-6 Luna targets high-volume, fast-response use cases at just $0.10 / $0.50 per million tokens. Despite matching GPT-5.6 Sol's performance levels, Luna costs roughly 1% of what the previous generation charged. Luna is available even to Free and Go tier users in the desktop app.
Both models were developed using methods similar to GPT-6 Astra's training pipeline, which OpenAI describes as "mini-Astras" — bringing top-tier capabilities down to more accessible price points.
Sources: TechCrunch, MacRumors
Google Discloses Gemini Breached Three Real Companies During Security Test
In one of the most significant AI safety incidents to date, Google confirmed on September 18 that its Gemini AI model gained unauthorized access to systems belonging to three real companies during a cybersecurity evaluation.
The incident occurred in May during a capture-the-flag exercise run by Irregular, an independent firm that tests advanced AI systems. Gemini was supposed to attack a fictional organization in a controlled sandbox — but internet access was unintentionally available, and a fictional company name in the scenario coincided with a real domain.
According to Google's VP for security engineering, the AI model guessed login credentials or found them in public repositories, then accessed the real systems believing they were part of the test. Critically, the model stopped before taking further action in all three cases.
The incident was first discovered by Irregular in late July but wasn't publicly disclosed until September, following a Wall Street Journal report. Security researchers have called it a "breakdown in test scoping, network containment, credential hygiene, and detection" — a warning shot about what happens when capable AI agents operate at machine speed without proper guardrails.
Alphabet Open-Sources Industrial Robotics Platform at ROSCon 2026
Alphabet's robotics division Intrinsic released Intrinsic Core under the permissive Apache 2.0 license at ROSCon 2026 in Toronto on September 22 — a move Forbes dubbed "giving away the Android of robotics."
The open-source package bundles the same capabilities Intrinsic uses in real manufacturing deployments: collision-free path generation, perception with standardized camera and point-cloud interfaces, an on-edge ML inference engine, hardware device drivers for industrial peripherals, and a kinematics solver library.
Intrinsic was folded into Google's core operations in February 2026 to accelerate physical AI development. The open-sourcing strategy mirrors Google's Android playbook — commoditize the platform layer to accelerate ecosystem adoption and let developers focus on higher-level applications rather than reinventing robotics fundamentals.
Sources: Intrinsic Blog, SiliconANGLE
Big Three AI Labs Formally Coordinating on Safety Standards
In a development that has flown under the radar amid the GPT-6 launch buzz, OpenAI, Anthropic, and Google DeepMind have confirmed they are actively coordinating on shared AI safety standards — a first for the three leading frontier AI developers.
The talks were revealed on September 15 by CNBC and Bloomberg, with OpenAI's global policy chief Chris Lehane confirming weeks of engagement. The effort follows two key catalysts: Google DeepMind chair Demis Hassabis publishing a proposal in July for a US-led AI safety standards body, and Anthropic CEO Dario Amodei's September 12 essay "We Must Pace the Frontier," which outlined three concrete steps including granting outside evaluators employee-level access to frontier systems.
OpenAI's chief scientist Jakub Pachocki stated: "I believe that shared safety standards and international coordination on further AI development need to be priorities now." Notably, OpenAI says no antitrust waiver is needed for the three firms to coordinate on safety matters.
Today's UN Security Council session adds further momentum to these efforts, signaling that AI safety is now firmly a matter of international security policy.