OpenAI publishes Block and SentinelOne Daybreak case studies

OpenAI on October 14 published two Daybreak case studies at once, featuring fintech company Block and security company SentinelOne. Daybreak is OpenAI's program that gives defenders access to its models for cybersecurity work. The common thread in both cases is time. SentinelOne cut malware reverse engineering that took a top researcher four to six weeks to about eight hours, and Block said that in recent red team exercises it took 18 minutes from submitting a high-level vulnerability to an automated proposed patch.

Block is the company behind Square, Cash App and Afterpay. According to OpenAI, Block's security team uses the OpenAI Defense Factory approach, in which AI handles vulnerability discovery, detection and response, and work that used to take weeks now takes hours. "AI is compressing the time from human intent to action," said James Nettesheim, Block's chief information security officer (CISO). "With things like Daybreak, defenders are able to fight fire with fire and keep up with the rapid pace of AI adoption."

Block's published figures center on detection work. Because every investigation stores its raw query evidence, the team can replay a proposed detection change against seven months of production telemetry in under an hour. That used to mean requerying hundreds of old cases by hand, so in practice it rarely happened. Since adopting automated triage, triage throughput is up 50%, and intelligence gathering and investigations are up 50%. The system performs about 1,100 analyst-hours of investigation per quarter, and about 600 analyst-hours per quarter of false-positive triage have been handed back. "Our system works every alert like it's the only alert we've got that day," Nettesheim said.

How the team runs models has also changed. Block's security team no longer defaults to the latest model and instead picks different models for different tasks. A security harness built on top of Daybreak routes tasks to the right model and includes hooks for humans to take over tasks or for agents to verify each other's work. The team keeps evaluating the system against internal benchmarks and real-world performance, and routes work to human review when that is the responsible path.

Two incidents show the effect. An alert on a single laptop first looked like a minor policy issue. 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. "One routine endpoint alert became a campaign takedown," Nettesheim said. "The signal only exists across two products' logs." In a recent red team exercise, the model was given a few hints about an internal system and chained together a full attack path of SQL injection, privilege escalation, a foothold, credential enumeration and a pivot to other systems in about two hours. A human red teamer would typically need several days.

SentinelOne's numbers come from reverse engineering. "Something that would've taken a top-flight researcher four or six weeks to do, AI is now doing at human-level output in something like eight hours," said Gregor Stewart, SentinelOne's chief AI officer (CAIO). SentinelOne's research teams used to reserve deep investigation for the most interesting or important threats and could reverse only very small numbers of pieces of software. With few samples to compare, there were few chances to spot links between threats. Faster analysis now lets researchers find a shared component in two apparently unrelated samples and use it as a lead in a customer investigation. Stewart called this "a kind of intelligence that we were never able to afford before."

The partnership turned into a product in September. According to SentinelOne's September 3 press release, which METAL reviewed, SentinelOne expanded its Wayfinder Frontier AI Services threat-exposure offering with OpenAI models delivered through the Daybreak Defense Network, starting with GPT-5.6-Cyber. SentinelLABS, the company's research arm, found in recent benchmarks that GPT-5.6-Cyber was best in class at reverse engineering and analyzing military-grade malware such as fast16. The new services come in two strands: AI-powered code risk analysis, which scans repositories for flaws, code implants, exposed secrets and supply-chain risk, and an AI-enabled compromise assessment that evaluates telemetry against detection rules. Both are rolling out in private preview. "Attackers are increasingly using AI to find and exploit weaknesses with greater speed and scale," said Steve Stone, SentinelOne's chief customer officer (CCO).

Seen through an AI engineer's lens, what both cases invest in is not raw model performance but where people are anchored in the loop. "The only thing that can't be automated is accountability," Stewart said, adding that decisions that change a system's architecture still need human eyes while point fixes are better suited for automation. Block likewise has people guide every finding through the same sequence, locating the code and its owner, proposing a patch, confirming exploitability, verifying the fix and deploying the change. Block's view is that agents do not carry an organization's context or the reasoning behind past decisions. As AI's judgment is proven in practice, Nettesheim said, "we'll slowly ladder up the risk level to use it for more sensitive systems."

METAL has previously reported that OpenAI published a Daybreak case study with Sophos, and covered the Defense Factory approach in which OpenAI put 250 people on an internal vulnerability sweep. With these two cases, Daybreak now reaches from managed detection and response to a payments company's in-house security team and a malware research group. The next task is to keep vulnerabilities from appearing in the first place. Nettesheim said he wants his team to spend less time patching and more on preventive, strategic work, and Stewart said he wants to bring OpenAI's cyber capabilities into code creation so developers can fix security issues as they build.