OpenAI published a customer story about Block on October 14, 2026, and the number it leads with is 18 minutes: the time the security team at the company behind Square, Cash App and Afterpay now takes to go from a submitted high-level vulnerability to an automated proposed patch. That class of work used to take days or weeks, the page says.
What was actually released
No new model and no new tool are announced. OpenAI Daybreak is the program Block’s security team already runs, presented here as the OpenAI Defense Factory at work, and the page describes how it is wired into vulnerability discovery, detection and response rather than what it costs.
Detection tuning now replays. Every investigation stores its raw query evidence, so a proposed detection change can be run against seven months of production telemetry in under an hour. That used to mean requerying hundreds of old cases by hand, which the page says rarely happened because it was slow.
Red teaming is compressed. In one exercise the model was given a few hints about an internal system and chained SQL injection, privilege escalation, a foothold, credential enumeration and a pivot to other systems in about two hours, against several days for a person. The page says the same chaining is how it surfaces vulnerabilities that isolated scanning misses.
Model choice became a per-task decision. Block’s harness sits on top of Daybreak and routes each job to a model it picks, with hooks for a person to take over and for agents to check each other’s work. Block’s chief information security officer, James Nettesheim, says the team used to reach for the newest model and no longer does.
What a buyer can take from this
- The replay is a data decision, not a model decision. Storing raw query evidence with every investigation is what makes a seven-month replay possible, and the pattern carries no vendor lock.
- Routing per task, with agents verifying each other, is the architecture detail here, and the one that survives a change of supplier.
| Figure as published | What it measures | What it is not |
|---|---|---|
| 18 minutes from a high-level vulnerability to a proposed patch | speed in recent red team exercises | not a deployed fix |
| Under one hour to replay a detection change against seven months of telemetry | a check the team can now run at all | not a saved cost |
| Roughly 600 analyst-hours a quarter handed back | time moved away from false-positive triage | not a headcount reduction |
| Approximately 1,100 analyst-hours of investigation a quarter | work completed | not work the team stopped doing |
| 50% more triage throughput, 50% more intelligence gathering | the customer’s own before and after | neither carries a published baseline |
One example shows the join the approach is sold on. An alert on a single laptop looked like a minor policy issue, and in peer review the model followed an indicator that appeared only incidentally in the evidence, matched it against Block’s identity logs, and surfaced a larger set of problems. Nettesheim says a routine endpoint alert became a campaign takedown: “The signal only exists across two products’ logs.”
What the claim rests on
Every figure on the page is stated by Block or by OpenAI. The page is a customer story, a sales document carrying the shape of a case study: nothing is audited, and no sample size, period or unit definition is given.
The 18 minutes carries the most caveats. It comes from exercises the page calls recent, with no count of how many, and it stops at a proposed patch. The page’s own account of the process has a person confirming exploitability, verifying the fix and deploying the code change, so speed at proposing is not speed at remediating. A team reading that figure as time to a fixed vulnerability is reading it one step too far.
The 600 analyst-hours are a reallocation, not a saving: that time used to go to triaging false positives and now goes to deeper investigations and more automation. The 1,100 hours measure investigations completed, which is output, not effort removed.
What is not published
No price, no plan, no availability date and no region list for Daybreak, which the page sells through contact sales. No baseline behind the two 50% figures, no period behind any of them, no sample size behind the 18 minutes, and no definition of an analyst-hour. The page tags the products Block uses as Codex, and the narrative never discusses Codex, so a reader cannot tell whether Daybreak runs on Codex, what it costs, or how the two relate.
On this page
What it changes for a security team buying agents
Two things here transfer to a team that will never buy from OpenAI. The first is the data model: keeping raw query evidence attached to every investigation turns a pile of closed cases into a test set you can replay a change against. That is a storage and schema decision, and it is why Block can answer in an hour a question that used to be answered by hand or not at all.
The second is that model choice is now a routing problem. A team that sends every task to its most capable model is paying for capability it does not use, and a harness that assigns work per task is the mechanism that fixes it. Block’s version also has agents verifying each other’s work, which is what makes an automated finding usable without a person redoing it.
The purchase stays opaque. Daybreak is not sold with a price on a page, so a buyer is comparing a contract against a contract, and the governance question comes first: who is allowed to deploy a patch an agent proposed. That boundary, not the model, is what a security leader signs off on.
The Neural Minute take
The figure to keep is not 18 minutes. It is where the 18 minutes stops.
We used to use the latest model, but you don’t need to do that anymore.
Nettesheim’s line describes a practice rather than a result: routing by task is a decision any team can make this quarter, with or without a supplier’s case study behind it. The 18 minutes is the supplier’s number, and it ends at a proposal.
Our reservation is about repetition as much as method. This is the second OpenAI Daybreak deployment we have read in a week, after the Sophos account published on October 9, and the two arrive in the same shape: a named customer, a headline percentage, an executive quote and no method. What Block adds is the routing detail and the replay pattern, which is why this one is worth a page and a second identical one would not be. None of the tools we document is Daybreak, so we have no verified price to set against it.
Three questions for a buyer before signing: how many exercises produced the 18 minutes; what the baseline and period behind the two 50% figures are; and whether the analyst-hours were removed from the team or moved inside it. The first two are arithmetic a supplier should have. The third decides whether the saving is real.
Sources
- How Block is building AI-powered security with OpenAI technology, OpenAI, published October 14, 2026: https://openai.com/index/block-ai-powered-security
- The 18-minute figure, the seven-month telemetry replay, the 600 analyst-hours, the 1,100 analyst-hours, the two 50% figures, the red team example and the Nettesheim quotes are read on that page on October 14, 2026, as Block’s and OpenAI’s own statements, not as independent measurements
- Our October 9 reading of the earlier OpenAI Daybreak customer story: https://www.neuralminute.com/news/sophos-cuts-threat-investigation-time-by-96-with-openai/