Here is a sentence I have been turning over since last week: OpenAI, the company whose models write a large share of the world's new bugs, just open-sourced a tool whose entire job is finding and fixing those bugs. The package is called Codex Security, it dropped as a public CLI and TypeScript SDK on GitHub, and the developer community responded the way it always does to something this loaded: 598 points and 230 comments on Hacker News in a day, roughly half of them about whether it works, and the other half about whether we should trust it.
My take, after reading the docs, the README, and the comments? It is the most logical move in agentic coding, and also the most awkward one. Let me explain why I think both things are true at once.
What OpenAI Actually Released
Before I get to the hot takes, the facts matter, so let me lay them out plainly.
Codex Security is not a toy. It is a command-line scanner that takes a repository you own, reads it with a frontier model, and produces findings that are sorted by severity, backed by evidence, and optionally patched. The default engine is gpt-5.6-sol running at maximum reasoning effort, which tells you OpenAI is not treating this as a side project.
Here is what stood out to me in the tool itself:
- It validates, it does not just flag. Scans are report-only by default. The tool separates real vulnerabilities from noise, tracks findings over time, and supports comparing scans to show which issues are new, persisting, or resolved.
- It can fix what it finds. Beyond detection, it attempts patches, and it integrates with Codex itself for interactive work, plus a containerized bulk-scan mode for big codebases.
- It is provider-agnostic. You can point it at OpenRouter, Fireworks, or Amazon Bedrock and run it on Claude, Qwen, or any other model. OpenAI shipping a security tool that officially supports rival models is a quiet admission that the market is now multi-model.
- It is built for CI. Environment API keys, headless output, SARIF exports, severity thresholds. This was designed to run in your pipeline from day one.
None of that is hype. The tool looks genuinely useful, and I say that as someone who is normally skeptical of security products with AI in the name.
The Part That Bothers Me
Here is where my opinion gets pointed. One of the top comments on Hacker News described AI security tools as "fire departments run by arsonists." It got a lot of upvotes, and not only from trolls. The concern is structural, not personal: the same models that accelerate feature development also accelerate the vulnerabilities inside it, and now the same company that sells the accelerant is selling the extinguisher.
I think that framing is half right. The other half is that the incentives are actually aligned. A coding agent that ships insecure code is a liability for its vendor, because every breach becomes a headline about that vendor's model. OpenAI has a genuine business reason to make its own ecosystem safer, and Codex Security is evidence of that. That is not nothing.
But there are real questions I do not have good answers to yet. Corporate teams will ask whether their source code should be uploaded to a cloud model for analysis, and the answer matters more than the marketing copy suggests. Security teams will ask what happens when the scanner says a codebase is clean and a vulnerability shows up in production anyway. Who owns that miss, the tool vendor or the engineer who trusted the tool? And startups like Snyk are watching closely, because a well-funded AI company shipping a scanner for free is exactly how a market gets disrupted.
My honest verdict: run it, but treat it like a brilliant junior security engineer, not a replacement for one. Let it scan, let it patch, and still have a human review the report before anything ships. The era of AI finding our bugs has arrived. The era of AI being solely responsible for our bugs should never arrive, and we are the ones who get to decide that.
Comments