Sophos says OpenAI Daybreak cut threat investigation from 38 minutes to 89 seconds. Here is what the vendor numbers actually measure.

The announcement and the numbers

On October 9, 2026, OpenAI published a case study describing how the security company Sophos has used models from OpenAI's Daybreak program to automate parts of its managed detection and response (MDR) work. The page's headline metrics are striking: the average response time for cases handled by AI agents fell from approximately 38 minutes to about 89 seconds, a reduction the two companies present as 96 percent, and 52 percent of MDR cases are now resolved end to end by AI.

These numbers are not the output of an independent audit. They are calculated and published by OpenAI and Sophos, the two parties selling the arrangement: OpenAI sells frontier model access through Daybreak, and Sophos sells MDR subscriptions to more than 625,000 organizations. That does not make the numbers false, but it does mean readers should understand exactly what is being measured, who chose the baseline, and which claims are documented and which remain unverifiable from the published material.

This article treats the vendor figures as reported claims, examines what they do and do not establish, and separates observed reporting from analysis and remaining open questions.

What the announcement says

The OpenAI page, dated October 9, 2026 and titled 'Sophos cuts threat investigation time by 96% with OpenAI Daybreak,' includes a 'Results at a glance' section with three headline figures: an average response time for cases using AI agents of 89 seconds, a 96 percent reduction in investigation time using OpenAI models, and 52 percent of MDR cases resolved end to end by AI.

The page describes the mechanics of the deployment. Sophos's Fusion system aggregates sensor data from more than 500 third-party integrations alongside Sophos's own products, generating what the companies describe as trillions of events every day. Sophos distills those events into roughly 1,000 to 2,000 daily cases for its nine security operations centres to investigate. An investigation agent assembles customer context, detections, indicators of compromise and relevant threat intelligence for each case, and a planning model runs a plan-execute-review loop that produces a summary with recommended response actions for analysts.

Sophos Chief Technology Officer John Peterson is quoted on the before-and-after comparison:

"Now, because of the agents we've been able to build through the Daybreak programme, the average response time for cases using those agents has fallen to about 89 seconds. About half of the cases we handle are now being automated by agents we developed using the Daybreak models."

Attribution: John Peterson, Chief Technology Officer, Sophos, quoted on OpenAI's case study page, October 9, 2026. The quotation proves what Peterson said; whether the figures it restates are accurate is assessed separately below.

What the numbers actually measure

The most important qualifier is easy to miss. The 89-second figure is the 'average response time for cases using AI agents.' It does not measure all Sophos cases. Cases not handled by agents, including anything routed to human judgement, are excluded from that average. So the correct reading is: among the subset of cases where Sophos let its Daybreak-built agents run, the agents finished in about 89 seconds on average.

The 96 percent reduction is derived by comparing that agent-handled average against a baseline that Sophos itself selected: its pre-Daybreak process, which the page describes as averaging around 38 minutes. Sophos chose that baseline, and OpenAI's page reports Peterson's own claim that the old performance was 'better than 96% of professional security operations centres.' If Sophos's human process was already unusually fast, the headline reduction overstates what a typical organization could expect; if it was representative, the reduction may be more transferable. The published page does not provide the methodology, sample sizes, date ranges or case-mix breakdown that would let an outside party check either way.

The 52 percent figure is also bounded. The page says AI resolves 52 percent of MDR cases 'within boundaries calibrated by Sophos analysts.' That means roughly half of MDR cases still are not fully automated, and the automated half was selected by criteria Sophos defined. Neither company published those criteria.

The baseline was already exceptional

One notable detail is that the page says Sophos's prior 38-minute process was, in Peterson's framing, better than 96 percent of professional security operations centres. If that self-assessment is accurate, the pre-agents baseline was already strong, which cuts both ways for interpretation: it makes the reported speedup more meaningful against a good process, but it also means most organizations without Sophos's scale and sensor coverage may not replicate it. This is Sophos's characterization of its own past performance, not an independently verified industry benchmark.

Three modes, three levels of customer control

A separate, arguably more consequential part of the announcement concerns customer control. Sophos has built three operating modes into its MDR service: Notify, Collaborate and Authorise. Under Notify, Sophos investigates and recommends but the customer acts. Under Collaborate, Sophos and the customer act together. Under Authorise, Sophos can respond directly on the customer's behalf.

According to the page, 'The same boundaries apply whether work is completed by a person or an agent,' and Peterson is quoted confirming the escalation rule:

"Anything we don't feel comfortable with an agent handling gets passed off for human judgement."

Attribution: John Peterson, Chief Technology Officer, Sophos, quoted on OpenAI's case study page, October 9, 2026.

For customers, the practical access question is which mode they operate in and how much autonomy they have granted. The page states that potentially destructive actions still require the appropriate level of human oversight, but it does not define which specific actions count as destructive, how mode switching works, or what an agent may do under Authorise without further approval. Those definitions matter more to risk than the speed metrics, and they are not in the published material.

What Daybreak is, and who gets it

OpenAI's Daybreak program page describes the offering as 'a governed cyber defense stack' combining frontier models, the Codex harness, Codex Security, trusted workflows and ecosystem partners. It outlines an agentic defense loop of inventory, discovery, dynamic validation, ownership assignment and verified remediation, with 'People review consequential changes and independently verify deployed fixes.'

Through 'Daybreak Access,' verified defenders can apply for more capable and permissive defensive tools paired with stronger verification, scope controls and oversight. OpenAI also describes a commitment of $1 billion in subsidized Daybreak access over six months for state and local governments, critical-infrastructure operators, community banks, nonprofits and open-source maintainers, and a Patch the Planet effort built with Trail of Bits that reports 661 patches produced and 458 accepted upstream across 65 codebases under review.

The key structural fact for enterprise buyers: Daybreak access is allocated by OpenAI. Sophos's automation gains depend on continued access to models that OpenAI gates through an application and verification process. A customer cannot buy Daybreak-grade capability on the open market the way it buys commodity software; it arrives through OpenAI's partner pipeline. This is a fact about the market structure, not a criticism, but it shapes who can achieve comparable results.

Analysis: what is demonstrated, and what is not

The following points are opinion and analysis, clearly separated from the reported facts above.

The Sophos result is credible as a demonstration that frontier-model agents can materially speed up the investigation stage of security operations at scale. The pipeline description is specific and internally coherent, and the scale, roughly 1,000 to 2,000 daily cases across nine SOCs, is large enough that a headline result built on nothing would be surprising. The report also makes a defensible economic argument: the page says the approach 'Helps Sophos scale compute rather than relying on equivalent growth in scarce cybersecurity headcount.'

At the same time, the published evidence has real limits. There is no independent audit of the timing methodology. The 96 percent claim has no published calculation. The selection criteria for agent-handled cases and for the automated half of MDR work are not disclosed. Error rates, false-positive handling, and whether agent-generated summaries changed the quality of analyst decisions are not reported. Speed without accuracy data is only half a productivity story.

There is also an unavoidable conflict of interest to weigh. OpenAI is marketing Daybreak through customer successes, and Sophos benefits from being showcased as a flagship deployment. Neither party has an incentive to publish unfavorable numbers. That is normal in vendor case studies across the software industry, but it means the appropriate stance is neither credulous dismissal nor acceptance: treat the figures as plausible vendor-reported results awaiting independent verification.

For security leaders, the most transferable takeaway may be the least flashy one. Peterson is quoted advising that the 'one thing security leaders should do tomorrow' is to refocus on fundamentals including patching, multifactor authentication, network segmentation and strong security operations, adding that 'Vulnerabilities are being discovered at an alarming rate and exploited at a scale that we've never seen.' That advice stands or falls independently of any Daybreak metric.

Questions that remain open

Several questions would make the result verifiable rather than merely impressive. What is the case-level definition of 'resolved end to end'? What percentage of agent-handled cases required human correction, and how is correction time counted? What is the accuracy and containment outcome distribution for agent-handled versus human-handled cases? Will either company publish the timing methodology or submit it for third-party review? And what happens to customer-facing guarantees if OpenAI changes model availability, pricing or access terms?

Until those are answered, the honest summary is this: Sophos reports a large, specific automation result from a governed frontier-model program, the result is consistent with what the underlying technology can plausibly do, and no outside party has yet been able to check the arithmetic. Readers should hold the 89-second and 52 percent figures as vendor-reported claims, not established facts, while noting that the direction of the reported change is far more likely than its exact magnitude to survive independent scrutiny.