OpenAI Touts GPT-6 Astra as Its Safest Model, But It's Still Dangerous

Details

OpenAI has released GPT-6 Astra with safety improvements described as addressing persistent problems in cybersecurity and prompt security. The AI Business report characterizes the announcement as a tension: OpenAI claims Astra is its safest model, yet the publication judges it "still dangerous" — a framing that reflects real residual risk rather than editorial hyperbole. The source content is extremely limited, providing no benchmark figures, failure rates, or comparisons to predecessor models. Based on the title and source framing, the article appears to cover the same system card release covered in more technical detail by other outlets: GPT-6 Astra's indirect prompt injection vulnerability (reportedly 8.5% failure rate in third-party evaluations), jailbreak resistance under adaptive multi-round attacks, and hallucination improvements over GPT-5.6 Sol. The "still dangerous" assessment is significant for enterprise buyers: it signals that even OpenAI's most safety-hardened release cannot yet be treated as safe for fully autonomous agentic deployments handling untrusted data. Cybersecurity teams evaluating Astra for sensitive workflows should seek out the full system card and independent benchmark data before deployment. This source provided minimal extractable content — the summary reflects the title and framing rather than full article analysis. Readers seeking technical specifics should consult OpenAI's system card directly.