OpenAI Touts GPT-6 Astra as Its Safest Model, But It's Still Dangerous
OpenAI's GPT-6 Astra claims 'safest' — but researchers already found exploitable gaps in cybersecurity defenses.
Why it matters
A new frontier model release with explicit safety improvements addresses a known weakness, but vulnerabilities remain. Practitioners evaluating GPT-6 Astra need to understand what 'safer' actually means before deploying it in sensitive environments.
The key facts
9 to knowModel: GPT-6 Astra (OpenAI)
Safety focus: cybersecurity vulnerabilities and attack resistance
Claim: 'safest model' released by OpenAI
Finding: safety improvements do not eliminate all dangerous behaviors
Published: September 4, 2026
GPT-6 Astra announced with safety improvements
Cybersecurity focus area for safety hardening
Model positioned as 'safest' but acknowledged as still carrying risks
Safety/capability tradeoff tensions emerging at frontier scale
Go to the source
AI Businessaibusiness.com
Publisher excerpt: With the new model. OpenAI addressed persistent safety problems, with a focus on cybersecurity.