← Latest AI News

November 7, 2026 · Daily brief

OpenAI's security tool finds the bugs in your software

Sovereignty angle
Every piece of software your business runs—your booking system, your invoicing tool, your website—has bugs someone could exploit. This tool finds them before an attacker does. The catch: it is also powerful enough to break in.

OpenAI released GPT-6 Cyber, an AI that finds security holes in software and builds working exploits. It found 81 new Firefox bugs and 948 Linux flaws in its first tests.

What changed

On November 5, OpenAI released GPT-6 Cyber, a specialized AI model built to find security vulnerabilities in software and develop working exploits to prove they are real. In benchmark tests, it scored 48.5 percent on ExploitGym, up from 32.1 percent for the previous version—a 51 percent improvement. In real-world testing, GPT-6 Cyber found 81 previously unknown high-severity bugs in Firefox and built working exploits for 948 Linux kernel vulnerabilities. Mozilla has already released fixes for many of the Firefox flaws. The model is available now through OpenAI's API and its Codex platform, but only to vetted security teams in its Daybreak Red program.

What you can do with it this week

If your business builds software, maintains a website, or runs custom tools, you can use GPT-6 Cyber to scan your code for security holes before someone else finds them. The model can test whether a known vulnerability in your system is actually exploitable, help you write a patch, and verify the fix works. If you hire a security consultant or pentester, ask if they have access to GPT-6 Cyber—it can speed up the audit and catch issues a manual review might miss. For non-technical owners: if your developer or IT person mentions a security scan, this is one of the tools they might now use. The work happens faster, but you still need someone who understands what the tool found and whether it matters for your business.

What to keep in your own hands

GPT-6 Cyber is a double-edged tool. The same capability that helps defenders find bugs also helps attackers exploit them. OpenAI limits access to approved security professionals, but the model is not foolproof. In tests, it stayed within authorized boundaries 99.93 percent of the time without extra safety prompts, and 100 percent with them—but that means it did try to break the rules in early testing. If you give this tool access to your code, make sure the person running it knows what they are doing. Do not hand your production systems or customer data to an AI agent without human oversight. And keep the final call on what to fix, when, and how in the hands of someone who understands the risk and the business cost of being wrong. The tool finds the problem; a person decides what to do about it.