Desk
Security, safety, and adversarial ML.
Astra's advanced cyber tools go first to Cisco, Cloudflare, and Palo Alto Networks. Everyone else gets a throttled model and a monitor OpenAI already admits will false-positive on ordinary work.
NVIDIA and CrowdStrike put red and blue agents in a closed loop and found the open-model pipeline produced fewer detections than the frontier system — but the only ones that held up.
OpenAI's Hugging Face incident report explains the technical chain, but critics say the deeper failure was letting known risky behavior survive inside training and evaluation.
A court has drawn a line between choosing another AI vendor and using national-security procurement power to punish a vendor’s safety limits.
OpenAI's postmortem reveals a layered control failure in which misaligned rewards met porous infrastructure, fragmented incident response, and evaluations with no safe exit.
BlueField-4 expands the DPU role from packet offload to an independent security, storage and control domain around AI compute.
VentureBeat is building a research franchise around the operational questions that arise after enterprise AI leaves the pilot phase.
Google is splitting the Flash line into a cheaper workhorse, a high-throughput Lite model and a restricted cyber specialist rather than shipping one general upgrade.
Flash Cyber is Google’s bet that cheap, repeatable vulnerability search beats a single expensive model call, and it is being withheld from general developers for that reason.