Block says AI frees 600 security analyst-hours a quarter, with human review retained

Block says AI has freed roughly 600 analyst-hours a quarter previously spent sorting false alarms, giving its security team more time for deeper investigations. In an October 14 customer case study published by OpenAI, the company described systems that investigate alerts and propose software fixes, while people still validate results and guide and deploy code changes.

From manual queries to reusable evidence

The starting problem was work that took too long to repeat. Reviewing a software bug or parsing a log could occupy days or weeks. Testing a change to a detection rule meant manually rerunning queries across hundreds of old cases. Because that process was so time-consuming, Block says it often did not happen.

Block now stores the raw query evidence from every investigation. That lets the team replay a proposed detection change against seven months of production telemetry—the records of activity in its live systems—in under an hour. The change makes a previously cumbersome check practical, rather than merely speeding up a task analysts already performed routinely.

The team also moved away from simply choosing the latest model for every job. Its security harness—the software coordinating the work—is built on OpenAI Daybreak and routes tasks to different models. It includes ways for people to take over and for AI agents to check one another’s work.

An alert becomes a wider investigation

Block reports a 50% increase in triage throughput, meaning the volume of alerts processed, and a separate 50% increase in intelligence gathering and investigations. Chief Information Security Officer James Nettesheim puts the automated investigation work at roughly 1,100 analyst-hours per quarter, saying every alert now receives a deep investigation. These are company-reported results presented in OpenAI’s customer story, not an independent evaluation.

One example began with a laptop alert that appeared to be a minor policy issue. During peer review, the model picked up an incidental indicator in the evidence and connected it with Block’s identity logs. Nettesheim says that connection exposed a much larger set of issues and turned a routine endpoint alert into a campaign takedown.

In a red-team exercise—a simulated attack used to test defenses—the model received a few hints about an internal system. Block says it assembled a full attack path in about two hours, progressing from a database vulnerability to higher privileges, access to credentials and movement into other systems. Nettesheim says that work would typically take a human tester several days.

The 18-minute patch figure also comes from recent red-team exercises. It measures the time to an automated proposed patch, not a deployed repair. Block’s remediation workflow includes locating the affected code, identifying its owner, confirming that the weakness can be exploited and checking that the proposed fix actually resolves it.

More sensitive systems remain the next step

Nettesheim draws a distinction between technical capability and organizational judgment. Agents do not carry the company’s context or the reasons behind earlier decisions, he says. People supply that context and strategy. Block continually assesses the systems against internal benchmarks and real-world performance, adjusting its approach and routing work to human review when needed.

As AI’s judgment proves itself in practice, Nettesheim says Block will gradually use it for more sensitive systems. That is a planned expansion of responsibility, rather than a claim that the current results justify unrestricted access.

For now, Block says the time recovered from false-positive triage is going into deeper investigations and more automation. Nettesheim’s longer-term aim is to reduce the vulnerabilities introduced in the first place, so the team can spend less time patching and more time preventing problems. The disclosed gains concern today’s investigation and repair workflows; prevention is the next ambition.