This time, the "star" of the open‑source release is codex‑security — a CLI tool plus a TypeScript SDK.
OpenAI has finally gone "Open" again! Recently, OpenAI officially announced that it had quietly released the open‑source code security powerhouse — Codex Security CLI.
Earlier, someone on Hacker News spotted it first, sparking an instant explosion of discussion, and its GitHub star count surged to 1.2k. Seeing that it could no longer be hidden, OpenAI had to come forward and claim it.
A few days ago, Jensen Huang personally stepped in and publicly voiced his support for open‑source AI. Soon after, OpenAI signed on and joined the camp. Unexpectedly, the surprise came so quickly — a heavyweight open‑source masterpiece was served up in the blink of an eye. It has to be said: Jensen’s words still carry serious weight!
1. OpenAI’s code security tool is now open‑source
The "star" of this release is codex‑security — a CLI plus a TypeScript SDK. Its core function is very straightforward, delivering a set of "combo moves": automatically discovering vulnerabilities in codebases, verifying them, and fixing them.
It emphasizes out‑of‑the‑box usability — just three commands are enough to run the entire workflow: for CI runs, no login is required, just configure an OPENAI_API_KEY. The prerequisites are Node.js 22+ and Python 3.10+, plus access to Codex Security.
1.2 million commits, 792 critical vulnerabilities
Strictly speaking, Codex Security is not a "new species." It evolved from the private beta project Aardvark launched in October 2025, and was renamed and released as a research preview on March 6 this year. The most hardcore aspect of this tool lies in its positioning — an application security agent.
Codex Security truly goes deep into the lower layers, reads your code, and understands what your system is actually doing. Its workflow consists of three steps: first, it reads the entire repository and generates an editable threat model, figuring out what the project does and where it is most exposed; then, based on this context, it looks for vulnerabilities and ranks them by real‑world impact; finally, it throws suspicious issues into a sandbox and conducts real stress tests — anything that cannot be verified is not reported.
Judging by its track record, it is indeed quite capable. In its first 30 days, it scanned over 1.2 million commits, extracting 792 critical‑severity and 10,561 high‑severity findings. OpenAI also noted that repeated scans on the same set of repositories reduced the false‑positive rate by more than 50%.
2. The first early adopters burned through their bills
No matter how grand the official promises, they are no match for the "real‑world disaster stories" that developers immediately shared. On HN, a developer named gregwebs took a small repository for a test drive and directly posted a crushing terminal log. At 0m3s, preparation began; at 1m20s, scanning started; then it ran all the way to 52m47s, and finally printed a red line — "The repository HEAD changed during scanning, please start over." Not only did an hour go to waste, but this single run consumed half of his weekly quota for the Pro plan.
Another user, Quai, had it even worse. The scan hit account rate limiting right at the start, and the tool gave up after retrying for one minute. The tool did prompt that "partial results were retained," but he could not find a clear way to resume from them in the next scan. That failed run cost about $13.
Why is it so expensive? A look at the default configuration immediately reveals the answer. Codex Security calls gpt‑5.6‑sol by default and brutally cranks the "reasoning effort" up to extra‑high. Keep in mind that Sol is the most premium tier in the GPT‑5.6 family, with API pricing at a hefty $5/1M input tokens and $30/1M output tokens.
3. Jensen leads the way, OpenAI takes a step
Right after Jensen Huang publicly called for supporting open‑source AI, OpenAI served up Codex Security.
However, don’t read too much into this "open‑source" move just yet. OpenAI played it smart: the open‑sourced part is the application‑layer shell, while the model layer remains firmly in its own hands. Nevertheless, the Pandora’s box of AI agents taking over code security has already been opened. Now, all eyes are on how developers will remix this open‑source marvel — using the magic of the community to beat OpenAI at its own game.
