Summary: OpenAI put Codex Security into research preview on March 6 for ChatGPT Enterprise, Business, and Education customers. The application-security agent is meant to find vulnerabilities across codebases and help prepare fixes, aiming at the backlog that forms after discovery. During the preview, flagged findings may be false positives and assisted repairs still require human review rather than automatic deployment.
Codex Security entered research preview on March 6, according to an OpenAI announcement. It is an AI application-security agent, built to do two things at once: surface the vulnerabilities sitting in a codebase, then help get them fixed. Who it serves is stated just as plainly — ChatGPT Enterprise, Business, and Education customers.
In its coverage, The Hacker News homed in on one point: scale. This agent carries out code scanning at scale. For OpenAI, the launch is the newest step in pushing agent capability into vertical markets. Security makes an instructive opening move — budgets there are big, the pain never lets up, and the work comes in just the shape agents handle best.
From Writing Code to Securing It
Security is a vertical application of what Codex could already do as a coding agent. Hunting vulnerabilities calls for precisely the kind of work agents are good at: parsing the layout of an entire codebase, following data as it moves between files, seeing how an input turns into attack surface — all while staying alert across tens of thousands of lines so nothing slips by. What limits the human security engineer has never been a shortage of knowledge; it is scale and stamina, and those are the two resources AI never runs out of. The extension also makes sense on its own terms: at bottom, a vulnerability is a bug with an adversary attached. The same grasp that lets an agent change code safely also lets it reason out how that code could be misused — the threat model sits one abstraction above the implementation.
Where Security Work Actually Stalls: the Fix
“Finds and helps fix” — four short words carrying more weight than their length implies. Discovery-only tooling is already everywhere: scanners emit backlogs that stretch on and never reach zero, since closing out a single vulnerability means grasping its context, comparing possible approaches, and verifying the patch leaves everything else working. Security debt piles up at precisely this step. Writing assisted repair into the product positioning aims straight at the costliest stretch of working hours in the security process.
The Enterprise-First Logic
Restricting the research preview to Enterprise, Business, and Education has a straightforward explanation. A security scan has to reach a company’s core code, which turns contracts, permissions, and data boundaries into preconditions — and the paid enterprise plans are the only tier carrying those preconditions out of the box. Enterprises double as the only buyers who fund “reduced security risk” directly: because the cost of a breach can be counted out, the ROI math practically writes itself. Consumers and individual developers will follow in time; the first move goes where the need and the money already are.
What Changes for Security and Development Workflows
- Checking turns continuous rather than scheduled: vulnerability discovery no longer hangs off quarterly pentests and instead travels with the code as it advances
- Security teams shift up a level: away from combing code line by line, toward vetting what the AI flags and proposes to repair, plus the cases that genuinely demand judgment
- Human review stays on during the research preview: an AI-flagged vulnerability may be a false positive, and repair suggestions call for sign-off even more — “helps fix” does not mean “auto-fix”
That running negotiation between automation and oversight is exactly what the research-preview posture exists for; expect the guardrails to loosen only once the evidence has earned it.