OpenAI released GPT-6 Cyber on November 5, a model tuned for authorized security work: vulnerability research, patch development, exploit development inside sanctioned programs. It is not generally available. Access runs through the API and Codex to customers vetted under Daybreak Red, OpenAI's program for authorized security testing.
Early testing data credits the model with identifying 948 distinct vulnerabilities in the Linux kernel. That figure will be quoted as proof that AI replaces security engineers. It is not.
What the model actually does better
Conventional static analysis excels at pattern matching: known dangerous functions, known vulnerable dependencies, known injection sinks. It is weak at code that looks correct but can be driven into a bad state.
A language model reads intent. Given a function, it can ask what happens if an attacker controls this input, follows this path, and races that thread. Applied to a codebase the size of the Linux kernel, with its decades of accumulated assumptions, that is how you surface boundary conditions no human reviewer had budget to chase.
How to read 948
Not every finding is exploitable. How many are reachable, which versions are affected, whether local privilege is required: all of that needs human triage. Treating raw output as a conclusion burns weeks on unreachable paths.
False positives remain expensive. The bottleneck in security work is not discovery, it is adjudication. A tool that produces a thousand leads without deduplication and severity ranking makes the queue longer, not shorter.
And access is gated. Most teams cannot use this model today and should not put it in a procurement plan. The signal is directional: vendors are building vertical variants for high-risk domains and wrapping them in approval processes, because the same capability writes exploits as easily as it finds bugs.
What to do without access
Do not wait. General-purpose models plus engineering discipline capture a meaningful share of the value:
Run review as a routine, not a spot check: every commit, focused on authorization boundaries and dangerous data flows.
Build a case library from your own past incidents, then ask the model whether new code resembles a known failure. That is far more reliable than asking it to imagine attacks.
Keep authorization explicit. Scope, time window, rollback plan, and never against production.
Leave the final call to a person. A model produces candidates; it does not carry liability.
The broader point is that offensive and defensive capability ship together. Any organization deploying this class of tool needs its own internal approval and audit rules, not just an API key.
GPT-6 Cyber Found 948 Linux Kernel Flaws. Most Teams Cannot Use It
← Previous: Gemini 4 Carbon Ships With 1M Context and a 60-Second First Token
Next: →Elvis and Eminem Rights Holders Sue Suno for More Than $202M →