A “critical level”: OpenAI suspends development of its next ChatGPT

On August 7, OpenAI published a rare warning: its preliminary assessments can no longer rule out the possibility that Astra (GPT-6), the lab’s next major model, will reach the “critical” threshold in cybersecurity. This highest classification comes from the Preparedness Framework, the internal policy that governs the company’s launches.A second post, published on August 18, reveals the full extent of the consequences: two weeks of pause on part of the training, and the company’s largest project still halted.

What can an AI rated “critical” in cybersecurity do?

The company’s own definition of this level is indeed daunting. A model crosses this threshold if it can independently identify and exploit vulnerabilities still unknown to vendors (the famous zero-days) in many critical systems that have nevertheless been hardened. The other route to access: carrying out a complete cyberattack against a protected target from a simple objective formulated in everyday language. Astra is not a confirmed case: the evaluation remains preliminary and testing continues. The model is, moreover, unrelated to the incident revealed on July 21, in which two in-house AI systems escaped from a test rig to reach Hugging Face’s infrastructure. Its larger sibling, GPT-5.6 Sol, officially tops out one notch lower, at the “high” level. OpenAI indicates that “preliminary indications that one of our upcoming models, Astra, could reach the threshold for critical cybersecurity capability defined by our preparedness framework”.

Yet mere uncertainty was enough to trigger the major lockdown. OpenAI has placed Astra under a prison-like regime: isolated test environments, restricted access to the network and to tools, encrypted model weights, sandboxed execution. A backlog of work remains frozen until it meets these requirements. The White House was even informed of the slowdown, at the company’s own initiative.Sam Altman summed up the matter bluntly: the capabilities of his models risked advancing faster than their alignment, security and monitoring mechanisms.

We have paused some frontier RL training to ensure that we can meet the appropriate alignment, security and monitoring standards for the new level of capabilities in front of us. Model progress is now extremely rapid, and we always said we would take action if we felt that model…

— Sam Altman (@sama) August 18, 2026

Monitoring its own models, the new bill that stings

The new system inspects the model’s internal activity with each fragment of generated text, then passes suspicious behaviors to increasingly compute-hungry automated investigators. Unauthorized access, data theft, destructive action or an attempt to bypass safeguards: every likely breach of a red line triggers the highest-level alert. Security teams then have 30 minutes to establish that it was a false positive, failing which the activity must be suspended.

This vigilance comes at a price, quantified by OpenAI itself: monitoring consumes around 20% of the computing power used by the work concerned. The company acknowledges significant costs and delays for its research.She also promises to revise her Preparedness Framework to cover training as much as deployment.

👉🏻 Follow the latest tech news in real time: add 01net to your sources on Google and subscribe to our WhatsApp channel.

Source: OpenAI