October 14, 2026
Block is using AI systems, including OpenAI Daybreak, to dramatically compress vulnerability remediation times, freeing up its security team to focus on strategic work.

- Company Size: Enterprise
- Region: North America
- Industry: Finance, Technology
- Product: Codex
Detection tuning
< 1 hour
Replaying proposed detection changes against 7 months of telemetry data
Proposed patch
18 minutes
From filing a high-severity vulnerability to automatically generating a proposed patch
False-positive triage
~600 analyst hours
Saved each quarter on investigating false positives
Investigations
~1,100 analyst hours
Of investigation workload completed per quarter
At Block—the parent company behind Square, Cash App, and Afterpay—the security team is leveraging the OpenAI Defense Factory: utilizing AI to handle vulnerability discovery, detection, and response at a scale human teams cannot match. Powered by AI systems including OpenAI Daybreak, Block accomplishes in mere hours what previously took weeks.
“AI is compressing the time between human intent and action. For a long time, attackers were widely seen as having the upper hand because they were faster and more agile. But with tools like Daybreak and the technical advancements across the OpenAI toolchain, defenders can now fight fire with fire and keep pace with the rapid proliferation of AI.”
— James Nettesheim, Chief Information Security Officer (CISO), Block
Block’s security team has been dedicated to responsibly integrating AI across diverse workflows, spanning both code generation and security operations. “The core idea is having a brand-new teammate that scales alongside you,” says CISO James Nettesheim. “It’s distinct from traditional automation, which requires rigid, step-by-step execution. Now, you can present it with a highly bespoke, ambiguous problem and work through the reasoning together.”
Initiatives like reviewing software flaws or parsing logs used to take days or weeks. Today, the team completes these projects in a fraction of the time, while humans continue to validate every finding.
In detection rule tuning, for instance, every investigation preserves its original query evidence. As a result, the team can now replay proposed detection rule modifications against seven months of production telemetry in under an hour. In the past, this process required manually re-running queries across hundreds of historical cases—a task so time-consuming that it was rarely performed in practice.
The team has also become more discerning when matching specific models to different tasks. “We used to default to whatever model was newest, but that’s no longer necessary,” he explains. “Learning to choose the right model for the right task is a learning curve in its own right.” Block’s security harness, built on top of Daybreak, routes tasks to the most suitable model and includes built-in hooks for human intervention or agent-to-agent cross-verification. The team continually evaluates these systems against internal benchmarks and real-world performance, adjusting strategies as needed—including routing tasks to human review whenever responsible handling is required.
Automated triage has unlocked both faster execution and higher task throughput. Block’s security team has observed a 50% increase in triage throughput, a 50% expansion in intelligence gathering and investigation capacity, and a brand-new autonomous threat-hunting capability. As Nettesheim puts it: “That equates to roughly 1,100 analyst hours of investigative work completed every quarter, because every single alert now receives a comprehensive, deep-dive investigation covering every component involved. Our system treats every alert as if it were the only one received that day.”
The OpenAI Defense Factory in Action
Traditional scanning tools often flag isolated vulnerabilities in silos, but Block’s security team goes much further with Daybreak. The model helps infer attack patterns and chain discrete attacks into end-to-end compromise paths, unearthing vulnerabilities that may have lingered undetected for years.
In one scenario, an alert on an individual laptop appeared on the surface to be a minor policy violation. However, during the review process, the model tracked an incidental indicator found within the evidence, correlated it with Block’s identity logs, and uncovered a far more extensive issue. “A routine endpoint alert ultimately turned into a full-scale campaign remediation. No single tool could have connected those dots, because the signal only existed across logs from two different products,” James noted.
Another showcase of Daybreak’s capabilities came during a recent red-team exercise. Provided with only minimal prompts and clues about internal systems, the model stitched together a complete attack path in approximately two hours—spanning SQL injection, privilege escalation, establishing a foothold, credential enumeration, and lateral movement into other systems. For a human red team, this process typically takes several days. During recent red-team operations, the turnaround from filing a high-severity vulnerability to generating an automated patch proposal was just 18 minutes.
Finding a vulnerability is only as valuable as the ability to remediate it. Nettesheim walked through the workflow his team follows for every discovery: pinpointing the exact location in the code, identifying the code owner, and progressing straight to remediation. “Propose a patch, confirm exploitability, verify the issue is resolved, and keep a human in the loop to guide the process and deploy the code changes.”
Despite their impressive technical capabilities, agents lack organizational context and the historical rationale behind prior decisions. Humans provide this crucial context and strategy, which defines the exact role his team plays—serving as the human review layer. As the AI’s judgment continues to prove itself in practice, Nettesheim notes: “We will progressively climb the risk ladder, expanding its application into increasingly sensitive systems.”
Key Results at a Glance
- Replaying proposed detection changes against 7 months of telemetry takes under 1 hour.
- In recent red-team exercises, generating an automated patch proposal from a high-severity finding took just 18 minutes.
- Approximately 600 analyst hours saved each quarter on false-positive triage.
- Approximately 1,100 analyst hours of investigative workload completed each quarter.
- 50% increase in alert triage throughput.
- 50% increase in threat intelligence gathering and investigation capacity.
What’s Next
Detection and response efforts alone have freed up roughly 600 analyst hours per quarter that were previously consumed by false positives. Armed with this reclaimed time, the team can now dedicate more energy toward deeper investigations and responsibly building out additional automated capabilities.
As attack tempos accelerate and AI models continue to evolve, Nettesheim sees an opportunity to drive down vulnerability counts…