The Scientist

SANS @RISK reports that SGLang's multimodal runtime passes unauthenticated ZeroMQ messages straight into pickle.loads(). The flaw is CVE-2026-93088, rated CVSS 9.8. CSO separately reports that just selecting a model in Unsloth Studio could run Python from that model's repo. Both bugs sit on hosts that hold GPU access, weights and cloud tokens. Port-scan those hosts from off-host today to confirm nothing is reachable, then rotate tokens and tighten how models get loaded this week.

The Information's account of OpenAI's forensics shows the July distillation campaign peaked at 16,000 requests from 4,000+ accounts over July 24–25. That works out to about 4 requests per account. Per-account rate limits and abuse classifiers cannot see traffic that thin; only clustering across accounts can. If you serve any model, including an embedding or ranking endpoint, you are probably watching the wrong unit.

TLDR Data reports that adding a model-built semantic layer lifted LLM analytics accuracy from 39% to 91%. The errors that remained came from hidden business conventions, such as which accounts count as customers. That gap is likely larger than the spread between frontier models on the same task. Google's PageBreak and a proposal to have LLMs write linters point the same way: deterministic structure around the model sets its precision.

OpenAI's Ultrafast tier charges 6x on both input and output tokens but only speeds up output. By Simplifying AI's derivation, each hour of waiting you buy back costs about $54 plus $10.8 times your input:output ratio. Separately, The Information reports HBM sold out through next year and Nvidia systems up about 17%. Devansh puts the fall in price for a fixed capability tier at ~940x in three years, so your token mix and memory footprint now drive cost more than list price does.