Critical web security flaws spark global alarm as Google, OpenAI and Amazon respond

OpenAI announced plans to make its latest large language models available through Google Cloud's Model Garden service. This service provides access to over 200 neural networks from Google, partners, and the open-source community. The first model OpenAI will bring to Google Cloud is its flagship GPT-6 Astra, responsible for several AI-generated math discoveries and achieving a perfect score on ExploitBench, a benchmark measuring LLMs' ability to find and exploit cybersecurity flaws.

Astra includes safety optimizations for enterprise use, causing "fewer misaligned outcomes" than other frontier models evaluated by OpenAI. Google Cloud customers will also have access to GPT-6.1 Sol and GPT-6 Luna, scaled-down versions of Astra, which cost five and 100 times less than Astra, respectively. These models will be generally available within a few weeks and will be followed by a launch on Google's Confidential Compute portfolio and Antigravity and Gemini Agent products.

Confidential Compute uses server-grade CPUs with a security feature called TEE, encrypting a segment of the host server's memory to secure customer workloads. Antigravity is a vibe coding tool, and Gemini Agent automates business tasks spanning multiple cloud applications.