OpenAI says Astra reaches its “critical” hacking threshold

In short: OpenAI says its upcoming Astra AI model can find and use previously unknown software flaws, so it will tightly limit who gets its strongest cyber features.

What happened

OpenAI announced that Astra, a future AI model, is the first one at the company to meet its threshold for “critical” cybersecurity abilities. OpenAI defines this as being able to independently find and exploit previously unknown weaknesses in real software.

The company said it followed its internal rules for this situation. It paused some development work for several weeks while it added new safety and security controls, then resumed work.

OpenAI plans to release a version of Astra to the public “soon.” But at launch, Astra’s more advanced cyber abilities will be available only to select partners in its Daybreak Blue early-access program, including companies like Cisco, Cloudflare, and Palo Alto Networks. OpenAI said the idea is to give defenders time to strengthen their systems before similar capabilities are widely available.

OpenAI also described new safeguards for everyday users, including a “misalignment monitor,” which is meant to detect and stop requests that look like hacking help. (Think of it like a security guard that watches for suspicious behavior.) OpenAI warned this monitor could sometimes slow down or stop legitimate activity by mistake, and users may be asked to review actions before continuing.

Why it matters

Tools that can discover new digital “unlocked doors” in software can help security teams fix problems faster. But the same skills can also be used for break-ins. OpenAI’s decision to limit Astra’s strongest cyber features shows how AI companies are trying to balance useful security work with the risk of misuse.

Source: Wired