OpenAI case study: how Block compresses security investigations and patch proposals to minutes

OpenAI published a customer case study on its site describing how Block builds AI-powered security with OpenAI technology. The subject is not a new model but a security pipeline that already runs in production.

Block runs Square, Cash App and Afterpay, and its security organisation covers vulnerability discovery, detection and response. According to the case study, putting AI inside those workflows changed the unit of time: investigations and log analysis that used to take days or weeks now finish in a fraction of that.

The numbers are stated plainly: a 50% boost to triage throughput, a 50% increase in intelligence gathering and investigations, and an entirely new autonomous hunting capability. The team says the system performs roughly 1,100 analyst-hours of investigation per quarter and hands back about 600 analyst-hours of false-positive triage. In a recent red-team exercise, the time from submitting a high-level vulnerability to an automated proposed patch was 18 minutes, and replaying a proposed detection change against seven months of production telemetry took under an hour.

The case study also describes a subtler episode: an alert on a single laptop looked like a minor policy issue on its own, but 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 issues. Block Chief Information Security Officer James Nettesheim says that signal only exists across two products' logs, so no single tool could have made the join.

Just as notable is where humans sit. The case study stresses that people remain in the loop: humans confirm exploitability, verify the fix and deploy code changes, while the model hypothesises about attack paths, chains them together and routes work between tasks. The team built a routing layer that assigns work to different models and keeps hooks for human takeover or for agents to check each other's work.

Nettesheim frames this as an asymmetry problem. AI is compressing the time from human intent to action, and defenders can now fight fire with fire instead of conceding an attacker advantage. He also concedes that agents do not carry an organisation's context or the reasoning behind past decisions, so humans supply that, and the risk level will be raised gradually as the AI's judgement is proven in practice.

What makes the case study useful is that it turns AI security from a slogan into operational metrics that can be checked. For security teams, the questions are whether such tooling plugs into existing alerting and logging, whether it preserves an evidence trail and a rollback path, and how far it is allowed to go in proposing patches automatically.

What to watch next is whether OpenAI packages the Defense Factory and Daybreak into offerings for a wider set of enterprises, and what standards the industry applies to auto-proposed patches once the capability reaches critical infrastructure and smaller organisations.