OpenAI Rated Its Own Model 'Critical' for Cyber Risk. Gate Your Agent.
Book: AI That Acts The series: AI in TypeScript — 5 books, from your first LLM call to agents in production — all five here My project: Hermes IDE | GitHub — an IDE for developers who ship with Claude Code and other AI coding tools Me: xgabriel.com | GitHub A customer uploads a PDF to your support agent. Page two carries a paragraph in eight-point grey that the human reviewer would never read,…
On 3 September 2026, OpenAI released its latest model, GPT-6 Astra, which the company rated as Critical for cybersecurity risk. The announcement emphasizes that the model can find previously unknown security flaws and develop new ways to exploit them across many well-protected systems without human guidance. The release includes new safeguards such as misalignment monitoring, new blocking alignment evaluations, and restricted deployment access.
These measures were implemented due to the model's ability to chain unknown flaws without human intervention. OpenAI co-founder and president, Greg Brockman, expressed his belief that GPT-6 Astra marks the beginning of the AGI era, although this is not a formal measurement. The system card provides information about the model's performance on various tasks, including its prompt-injection attack success rate of 8.5%, which is an improvement from its predecessor, Sol, with a rate of 27.0%.
Despite this reduction in attack success, the number still represents a real threat, particularly when considering the various paths in which attacker-controlled text can reach the context of an agent.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
