[DEVELOPING] OpenAI curbs Astra training on 'critical' cyber capability as Microsoft patches a one-click Copilot flaw
OpenAI is reported to have imposed stricter controls on Astra model development and halted many training workloads after testing indicated the model could reach 'critical' cybersecurity capability, while Microsoft shipped a patch for CoSnitch, a one-click Copilot data-exfiltration vulnerability it had known about for roughly eight months. The pair puts frontier-model offensive capability and shipped-agent security on the same day's tape. The desk's cyber print calendar is close behind: CrowdStrike and Marvell on 08-27, Palo Alto Networks on 09-01.