Block cut a proposed security patch to 18 minutes. Humans still control deployment

According to a case study published by OpenAI, Block uses systems including OpenAI Daybreak to investigate security alerts, connect evidence and prepare fixes. In recent red team exercises, 18 minutes elapsed between submitting a high-level vulnerability and receiving an automated patch proposal. That was not an automatic production deployment.

Seven months of telemetry fit into an hour-long test

The team stores the original queries and evidence for every investigation. It can therefore replay a proposed detection change across seven months of production telemetry in under an hour. Previously, analysts had to rerun queries manually across hundreds of old cases, so the company says this check often did not happen at all.

Block reports a 50% increase in triage throughput and a separate 50% increase in intelligence gathering and investigations. Automated investigations account for roughly 1,100 analyst-hours per quarter. Reducing manual false-positive triage has returned about another 600 analyst-hours per quarter.

The agent changes the cost of diligence, not ownership of the outcome

The more interesting shift is the economics of verification. If old cases can be replayed in an hour, a team need not reserve deep analysis for a small selection of alerts. The system can examine every alert more thoroughly and hand people connected evidence rather than a long queue of isolated signals.

For security teams, that creates a practical requirement: the agent needs retained evidence, clear code ownership and a path to human review. Without those supports it merely produces proposals faster. With them it can make previously skipped diligence affordable.

Customer-reported figures measure proposals, not deployed fixes

OpenAI published the results as a customer case study. It is not an independent benchmark, and the page provides neither raw data nor a method for calculating the reported analyst-hours. The 18-minute measure also ends at a proposed patch. Block's team still confirms exploitability, reviews the fix and has a person direct deployment.

The same distinction applies to a red team example in which the system assembled an attack chain in about two hours, from SQL injection through privilege escalation and movement into other systems. It is a strong internal result, not a universal measure of Daybreak's performance in another environment.

Regressions and reviewer interventions will settle the claim

The next useful figures are patch rejection rates, missed vulnerabilities and the frequency of human intervention before deployment. Only sustained results can show whether higher throughput improves defense or merely creates more work to review.

Transferability will matter too. Block is building on its own telemetry, internal benchmarks and a security harness that routes work among models. Copying one prompt will not reproduce that operation.

Lilith's verdict

An 18-minute patch is impressive only until a reviewer finds a flaw and refuses to let it into production. Block is right to keep the final finger over the deploy button human.