OpenAI AI Updates: August 8, 2026

1. OpenAI Flags Its Astra Model as Approaching the Highest Cyber Risk Tier

OpenAI. OpenAI published preliminary cybersecurity evaluations for a new model it calls Astra and said it slowed parts of the model’s development after internal testing showed capabilities strong enough that it could not rule out the maximum “critical” risk classification under its preparedness framework, a first for the company. The concern is that the model could independently identify and carry out cyberattacks, so OpenAI paused portions of the work while it strengthens safeguards and security controls before wider deployment. For practitioners, it is a rare public admission that a frontier lab held back a model on capability grounds, and a signal that offensive-security evaluations are becoming a gating step rather than an afterthought. Source

2. OpenAI Details HSP GRUPPE’s ChatGPT Enterprise Deployment for Tax Advisory

OpenAI. OpenAI published a case study on German firm HSP GRUPPE using ChatGPT Enterprise to build internal AI capabilities for its tax advisory practice, framing the rollout as a productivity and client-service play rather than a single tool. The write-up describes staff using the assistant to speed up research and drafting inside a regulated professional-services workflow. The example is a data point on how conservative, compliance-heavy fields are folding general-purpose assistants into day-to-day advisory work. Source