OpenAI's security agents needed context outside the code, so it is writing it down

A speaker on OpenAI's DevDay 2026 session on Codex Security said security agents need context that lives outside the code. Architecture docs, firewall settings and threat models were part of what OpenAI wrote down for them. The transcript does not name the speaker.

What was said

The speaker said OpenAI first pointed an agent scanner at its repositories with a broad order. The instruction, in their words, was "find all vulnerabilities, make no mistakes, didn't really cut it." So the team ran many scanners, each focused on one part of the code.

The speaker said each agent needed enough context to understand what it was looking at. That means knowing the purpose of the part, how it fits the whole, how it is exposed and which protections already surround it. Some of that context, they said, lives outside the code base.

OpenAI built two layers. The first is a custom instruction for each scan. It sets which areas to examine, what is in or out of scope, and which kinds of weakness to stress. The second is a shared knowledge base, reused across scans. It holds items such as architecture docs and firewall configuration. The speaker said it helps the agent grasp protections that cannot be seen in source code.

OpenAI also began committing security markdown files across components of its repository. The speaker compared them to the instruction files that coding agents read. They said this lets product owners define their own threat models and own their security guidelines.

Why it matters

Our reading: documentation used to be written for new hires and auditors. In this account it becomes working input for a scanner. If a firewall rule lives only in a console, the agent has no way to see it.

For engineering leaders, that puts the architecture page and the firewall export inside the security toolchain. The example also gives the writing to product owners, not to a central security team. Buyers of security tools can reasonably ask what context a product expects and who must keep it current.

The other side

The excerpt gives no measurement of how much the context improved results. We have only OpenAI's account that it was critical, and that the custom scan instruction helped a great deal.

The excerpt also does not say how OpenAI keeps these documents current. Whether a stale firewall description would mislead an agent, as a missing one might, was left open. And this is one company's experience with its own large code base. Smaller teams may find such documents harder to write and maintain. The speaker also noted that the agents read the code itself. The documents add to that reading rather than replace it.