OpenAI is handing the "first look" on part of its security team's workload to its own models: new alerts get triaged by AI first, and humans step in only for high-impact decisions. In a security-focused post published on August 17, Greg Brockman put the arrangement on the record, alongside a more specific argument: security teams shouldn't just use AI to find more vulnerabilities — they should use it to shorten the path to fixing the ones that actually matter.
This isn't a standalone product launch. It reads more like OpenAI's public account of how it's pacing the automated attack-and-defense cycle: models can already chain together vulnerabilities, misconfigurations and over-privileged identity relationships into full attack paths, and if defenders keep working through them at the speed of a traditional ticket queue, the backlog will outgrow their capabilities before those capabilities catch up.
Alert triage goes automated first
OpenAI says "almost all" initial security alerts today pass through an AI system for triage before a human is looped in. The company is also connecting detection results to bounded automated responses, while keeping high-impact decisions with people.
The post puts it directly: "The goal is to ensure we can detect and respond to security issues at machine speed." In plainer terms, the repetitive screening work that machines are suited for gets sped up, while privilege escalation, production changes and incident judgment calls stay with security staff.
The personal example in the post illustrates the workflow OpenAI wants to promote: Brockman had the publicly available GPT-5.6 Sol review his personal static site, which turned up 13 issues in about 15 minutes; the model then helped fix DNS, TLS, dependency and deployment issues over roughly the next hour. The example comes from the company's own leadership and shouldn't be read as a general performance benchmark — but it shows the direction of compressing "find, verify, fix" into a single workflow.
Scan the narrowest slice of the path first
OpenAI isn't recommending that companies go straight to building an unmanned security operations center. The path it lays out is deliberately conservative: start with a read-only scan of a single code repository, then have the model read through already-closed alerts; once that process proves reliable, move on to pull request review, live alert triage, and narrowly scoped automatic closing of false positives.
That sequencing is worth noting. Security automation most often goes wrong not because a model can't spot an anomaly, but because of what permissions that anomaly gets connected to. The post repeatedly stresses traditional controls — least privilege, network isolation, monitoring, staged rollouts. Once AI enters the workflow, those controls don't become obsolete; they become more necessary to keep explicit.
For engineering teams, the metric worth watching right now shouldn't just be "how many issues did the model find." More useful numbers are: how long it takes a high-priority issue to go from discovery to confirmation, whether fixes ship with regression tests, and whether automated actions stay within permission boundaries at every step. OpenAI's latest statement pushes those engineering metrics to the front of the AI-security product narrative.
Sources: OpenAI security post, CocoLoop, verification of AI alert triage claims, the human decision boundary and the personal-site case study.