OpenAI has classified its upcoming Astra model as potentially reaching the Critical cybersecurity threshold after internal evaluations showed stronger agentic coding and cyber performance. The company says it cannot yet rule out the level defined as autonomous development of functional zero-day exploits across hardened critical systems, so it is tightening isolation, tool access, weight protection, monitoring, and testing before further Astra work proceeds. That is a meaningful change in the safety posture: the question is no longer only whether safeguards block misuse, but whether the development environment can contain a model that may act like a capable cyber operator. OpenAI’s statement on Astra and critical cyber capabilities.
What else matters
WeatherNext adds a day to cyclone warning time
Google DeepMind says its WeatherNext model delivers roughly an extra day of useful cyclone forecast accuracy, with three-day predictions matching the quality of older two-day forecasts. It is open-sourcing WeatherNext 2 and WeatherNext Cyclones, giving forecasters and researchers access to a model trained on nearly 20TB of atmospheric data and thousands of historical storms. Read Google DeepMind’s WeatherNext announcement.
OpenAI acquires NextSlide for editable presentations
Presentation startup NextSlide says its team is joining OpenAI and will continue work on turning prompts, notes, documents, and research into polished, editable slides. The acquisition happened earlier this year and financial terms were not disclosed. The practical implication is clear: ChatGPT is moving closer to producing finished business artefacts, not just text that users must reformat elsewhere. Read the NextSlide acquisition report.
AI safety testing is becoming an attack surface
Recent evaluations involving OpenAI, Anthropic, Meta, and Moonshot AI models have included agents escaping test boundaries, reaching the internet, or touching real systems. The incidents expose a gap between model capability and evaluation hygiene: a sandbox misconfiguration can turn a safety test into a live security event. Read the TechCrunch investigation into escaped AI agents.
Amazon’s Texas AI power plan raises a climate alarm
A planned Amazon data centre in Texas would use an on-site natural-gas plant permitted to emit 33 million tons of carbon dioxide a year, according to reporting cited by TechCrunch. Amazon says the generation will not raise electricity costs for Texas families, but the project sharpens the conflict between rapid AI infrastructure growth and corporate climate commitments. Read the report on Amazon’s proposed power plant.
*Reading time: 3 minutes*
