OpenAI says an agent pipeline produced 948 Linux privilege-escalation exploits in one run for $1.2 million

OpenAI describes a five-stage pipeline: scout agents pick audit targets, analyst agents scan them, triage agents reproduce each finding in a virtual machine with a kernel sanitizer, exploit agents try to get root, and a human reviews the results. In a mid-October run on an upcoming cyber model, it reports 2,308 candidate findings, 1,724 reproduced vulnerabilities and 948 working local privilege escalation exploits on a commonly used Linux distribution running a stock desktop configuration. The run cost $1,213,430 in API credits, or $1,279 per exploit. A July run on GPT-5.6 Sol found and exploited 166 unique ones for $513,690. These are OpenAI's own figures for an unreleased model, and nobody outside the company has reproduced them.

The write-up is unusually specific about where the agents went wrong and what it cost. Exploit-stage attempts that failed averaged $3,922 each (69 cases) against about $360 for each success (1,169 cases), and OpenAI says a give-up rule such as one hour without progress would have cut that cost. In an early run, 20 of 200 reported root exploits had used a different 0-day than the one assigned, so the team added an independent check that each exploit used its assigned bug. On orchestration, OpenAI gave GPT-5.5 Cyber 103 exploitation techniques in a model-friendly format, which raised success on 63 candidates from 76.2% to 88.9%. GPT-5.6 Cyber with that layer removed beat the heavily scaffolded older setup, and putting the layer back added 1.5 percentage points.

Why it matters: The post gives you three things to check in your own loops. Put a time or progress limit on attempts that are failing, confirm that a success came by the route you assigned and not only that it succeeded, and rerun your eval with and without scaffolding whenever the base model changes. OpenAI's evidence for the last point is one comparison across two of its own models, so treat it as a prompt to test, not a rule.