OpenAI agents exploit 948 Linux kernel bugs for about $1.2M in credits OpenAI agents exploited 948 Linux kernel bugs, including working privilege-escalation exploits on a hardened Linux configuration, at a cost of about $1.2 million in credits — roughly $1,279 per working exploit. The result means kernel exploits that were once specialist work can now be produced by the hundred from an API budget, making the delay between a kernel fix and its deployment the deciding factor in exposure. That is about $1,279 per working privilege escalation exploit on a hardened Linux configuration. Kernel exploits were specialist work, and at this price they come by the hundred from an API budget, so the delay between a kernel fix and its deployment now decides exposure. Read: That is about $1,279 per working privilege escalation exploit on a hardened Linux configuration. Kernel exploits were specialist work, and at this price they come by the hundred from an API budget, so the delay between a kernel fix and its deployment now decides exposure. Read: We covered the $2/$10 API launch on Oct 24, when weights were promised by Nov 4. They are now on Hugging Face, and Vercel AI Gateway serves the model from several zero-retention providers, so teams can self-host it or avoid sending data to Moonshot. Read: Developers now choose between Flare, which returns low-quality images in as little as six seconds, and Sunburst, which holds edits consistent. Markup input and SVG export let an app edit a marked region or ship a scalable asset without a separate vector step. Read: Bloomberg kept most of each migration deterministic and gave the LLM only bounded file edits across thousands of repositories. Owners merged about 31% of the pull requests, a production yield teams can plan review load against instead of a benchmark score. Read: Adapted vec2text recovered 47.4% of short health instructions exactly from text-embedding-3-small and mistral-embed vectors. A RAG store that keeps only embeddings still holds readable private text, so its vectors need the same access controls as the documents. Read: Background subagents now get their 30-second status lines from Haiku 5.5 reading short notes, not the full conversation. Sessions use about 5% less of the plan limit, so the same subscription covers roughly 6% more prompts with the subagents' own model unchanged. Read: Unit volume rose about 65% from fiscal Q2 to Q3 2026, and the rest of the doubling came from more powerful chips. Non-Nvidia accelerator supply is growing faster than unit counts alone suggest, which adds capacity for the labs that train and serve on TPUs.