Block’s OpenAI-Powered Security Hunts Bugs in Hours, Not Weeks: Inside the Daybreak Harness
The security team at Block, the company behind Square, Cash App, and Afterpay, had a familiar problem. Reviewing a software bug or parsing a log file took days or weeks, and detection rules that should have been tested against old telemetry rarely were, because the work was too slow and too painful to repeat. According to OpenAI, which published a case study on the collaboration on October 14, 2026, the company solved it with an AI harness called Daybreak, built on OpenAI's Defense Factory program. Now the team does in hours the work that used to take weeks.
What Daybreak actually does
Daybreak is described as a model-routing harness. When a security task comes in, the system picks the OpenAI model best suited to that specific job instead of defaulting to the latest flagship release. Nettesheim said the team had to unlearn its old habit of always reaching for the newest model. Different models handle different kinds of security reasoning better, and the routing layer is what makes that choice automatic.
The harness carries two safety valves. At any point a human can take over a task, and separate agents can verify one another's work before a result is trusted. Block says the team continually tests the systems against internal benchmarks and real-world performance, and routes work to human review when that is the responsible path. The guardrails matter because the system operates on a payments company that moves real money for real people.
The numbers Block reports
Block's published results are unusually concrete, and they are worth quoting carefully because they come from the companies themselves rather than independent measurement. Triage throughput is up 50 percent. Intelligence gathering and investigations are also up 50 percent. The system performs roughly 1,100 analyst-hours of investigation per quarter, because, as Nettesheim said, every alert now gets a full, deep investigation across everything it involves.
Detection tuning shows the most dramatic change. Every investigation stores its raw query evidence, so the team can replay a proposed detection change against seven months of production telemetry in under an hour. The old way meant requerying hundreds of old cases by hand, work so tedious it rarely happened in practice. False-positive triage has handed back roughly 600 analyst-hours per quarter, time that used to burn on alerts that meant nothing.
In red-team exercises, Daybreak chained a full attack path, from SQL injection through privilege escalation to credential enumeration and lateral movement, in about two hours. A human red teamer would typically need several days for the same chain. In recent exercises, the time from submitting a high-level vulnerability to an automated proposed patch was 18 minutes.
One laptop alert became a campaign takedown
The most interesting example in the case study is not the headline speed number. It started with an alert on a single laptop that looked like a minor policy issue. During peer review, the model followed an indicator that appeared only incidentally in the evidence, joined it against Block's identity logs, and surfaced a much larger set of problems. Nettesheim called it a campaign takedown born from a routine endpoint alert, and his explanation is worth attention: no single tool could make that join, because the signal only existed across two products' logs.
That detail points to what the system really is. Traditional scanning tools flag one vulnerability in isolation. Daybreak is built to hypothesize about attack patterns and chain a series of attacks into a full end-to-end compromise path. It surfaces vulnerabilities that may have existed for years because nobody had a way to connect the dots between siloed logs.
A human still signs off every fix
Block's remediation workflow has not been handed over to the machines. Nettesheim walks through the same process for every finding: locate where the issue lives in the code, identify who owns it, propose a patch, confirm it is actually exploitable, and verify the fix. A human guides that chain and deploys the code changes.
The reasoning is straightforward. Agents can be technically capable, but they do not carry an organization's context or remember why earlier decisions were made. A human layer of review provides that strategy. As the systems prove their judgment in practice, Nettesheim said the team will slowly raise the risk level and trust the automation with more sensitive systems. The phrase is worth noting: trust is laddered up, not granted wholesale.
Why this matters beyond Block
The Defense Factory framing is OpenAI's broader bet. The pitch is that defensive AI capability, once limited to large technology companies and governments, can now reach critical infrastructure operators, schools, and small businesses. Whether that democratization materializes is still an open question, since the current disclosure centers on one sophisticated customer with a large security team.
Even with that caution, the shape of the result matters. The bottleneck in corporate security has never been a shortage of tools. It has been the shortage of attention: too many alerts, too little time, and detection rules that never get tested because testing them costs days. A system that gives a security team back 600 hours a quarter does not eliminate the human analyst. It lets the analyst do the work they were hired for.
Post Comment