Evidence at a glance
How an Ordinary Alert Became a Broader Investigation
On October 14, 2026, OpenAI published a case describing how Block uses AI for vulnerability discovery, detection, and response. Block is the company behind Square, Cash App, and Afterpay, and the account focuses on how its security team uses AI systems including Daybreak to handle investigations, rather than on a new consumer security product. One investigation began with what looked like a minor laptop policy alert.
Block says that during review, the model followed an indicator that appeared incidentally in the evidence and connected it with identity logs, uncovering a broader set of issues. The example matters not because the model independently “saw” an attack, but because clues were distributed across product logs and only became meaningful when joined. The operational bottleneck is therefore not just how long it takes to read an alert. It is also whether a team can assemble its available evidence into a path that can be investigated.
Retained Evidence Makes Historical Replay Practical
Block says each investigation retains its raw query evidence, allowing the team to replay a proposed detection change against seven months of production telemetry in less than an hour. Previously, similar checks required people to requery hundreds of old cases, a workload so expensive that it often did not happen. The change is not simply making one query faster. It is lowering the cost of repeatedly checking a rule.
Once replay is practical, a team has a chance to examine how a rule change behaves against historical evidence instead of relying only on a small number of new alerts. But this capability depends on evidence having been retained and remaining usable. If relevant logs are missing, query evidence cannot be reused, or systems cannot be connected, a model cannot supply information that was never recorded.
The Workflow Defines Both the Model’s Role and Accountability
Block’s security workflow is built on Daybreak and routes tasks to suitable models, with provisions for human takeover and agents checking one another’s work. The team also adjusts its approach using internal benchmarks and real-world performance. James Nettesheim, Block’s security chief, describes AI as a new teammate that people can work with. That framing does not imply fully autonomous operation: models can handle more ambiguous, tailored tasks, while people provide organizational context and decide when to intervene.
The patch figure especially needs to be read as one stage in a process. OpenAI reports that during a recent red-team exercise, it took 18 minutes from submitting a high-level vulnerability description to the model automatically proposing a patch. The case describes further steps: confirming exploitability, validating the fix, and having people guide and deploy the code change. So 18 minutes measures the time to a proposed patch, not the time to review, validate, and ship a completed fix.
The Efficiency Figures Do Not Add Up to One Savings Number
Block reports a 50% increase in alert-triage throughput and a 50% increase in intelligence gathering and investigations. The case also says that reducing false-positive triage returns about 600 analyst-hours per quarter, while the system completes roughly 1,100 analyst-hours of investigation in the same period. Those latter figures describe different things: 600 hours is time returned to the team, while 1,100 hours is investigation work performed by the system. The second figure should not be recast as labor saved.
A separate red-team example says Daybreak took about two hours to connect clues from a small number of internal systems into an attack path involving steps such as SQL injection, privilege escalation, and credential enumeration. OpenAI says human red-teamers would typically need days. This comparison conveys the speed difference the case aims to show, but the public account provides no exercise plan, control group, or independent review. The figures also lack complete measurement methods, sample sizes, and error ranges, so they should be treated as Block’s reported results relayed by OpenAI, not as industry benchmarks that can be applied directly.
Start with the Evidence Chain, Not the Model Purchase
The practical lesson is to identify security tasks that are skipped because they cost too much. Historical alert replay, deeper false-positive investigation, and cross-log correlation are possible candidates, but each task first needs a clear account of the evidence it requires, how long that evidence is retained, who reviews the result, and which outcomes must be handed back to a person. Model selection and task routing have meaningful boundaries only when they sit inside this arrangement of responsibilities.
The available material does not explain how Block authorizes model access to identity and endpoint logs. Nor does it provide an independent audit, full measurement definitions, or data on security outcomes after remediation. Other teams can borrow the workflow questions, not simply copy Block’s reported percentage gains: establish an internal baseline for one replayable investigation, retain the underlying evidence, and define human review and takeover conditions. If logs are unavailable, results cannot be verified, or no one owns the decision, shorter generation time alone does not demonstrate stronger defense.