Sophos Says AI Agents Cut Threat Investigations From 38 Minutes to 89 Seconds

Sophos says agents built through OpenAI's Daybreak program have pulled the average investigation time for managed detection and response cases down from roughly 38 minutes to about 89 seconds — a 96% reduction — and that 52% of its MDR cases are now resolved end-to-end by AI. The claims come from an OpenAI customer story published October 9, 2026, with figures attributed to Sophos Chief Technology Officer John Peterson. It is one of the more specific operational numbers any security vendor has put on agentic triage, and also one that no third party has verified.

What Sophos says it built

The work sits inside Sophos Fusion, which the company describes as its AI-native cyber defense system and which includes its Managed Detection and Response service. According to Sophos, Fusion ingests sensor data from more than 500 third-party integrations alongside the company's own products, generating trillions of events per day. Those events are distilled down to roughly 1,000 to 2,000 cases a day for nine security operations centers to work through — the funnel that matters, since the economics of MDR are set by how many human-hours each case consumes.

The agent architecture Sophos describes is conventional in shape and unremarkable in a good way. An investigation agent assembles the raw material for a case: customer context, detections, indicators of compromise and relevant threat intelligence. A planning model then runs what Sophos calls a plan–execute–review loop — it drafts an investigation plan, works through the steps, and produces a summary with recommended response actions for an analyst to review. Additional agents can execute parts of the response.

That pattern — retrieval, plan, execute, summarize, hand to a human — is close to what most serious agent deployments in regulated or high-consequence domains look like right now. The interesting part is less the design than the claim that it holds up against live attacker traffic across a stated base of more than 625,000 protected organizations.

The baseline matters more than the percentage

A 96% reduction sounds dramatic until you look at what it is measured against. Sophos's pre-Daybreak process averaged around 38 minutes per case, and Peterson says that figure was already better than 96% of professional security operations centers. In other words, the comparison is against a strong internal baseline, not an industry median — which cuts both ways. It makes the improvement harder to dismiss as low-hanging fruit, but it also means other MDR providers starting from a slower baseline should not assume the same multiple.

The 89-second figure also comes with a carve-out stated in the source: it is the average response time for cases using those agents. The 52% of cases closed end-to-end by AI and the 89-second average are describing overlapping but not identical populations. Sophos does not say how the remaining roughly half of cases are distributed in time, nor whether the agent-eligible cases skew toward simpler, more templated incident types. That would be the natural first question from anyone who has run a SOC: automation tends to land first on the high-volume, low-ambiguity tail, and the residual human workload is often the expensive part.

Human oversight is a product feature, not a disclaimer

Sophos frames its guardrails through three MDR operating modes that predate this work: Notify, where Sophos investigates and recommends but the customer acts; Collaborate, where both parties agree before action; and Authorise, where Sophos can act directly on the customer's behalf. The company says the same boundaries apply whether a human or an agent does the work, and that potentially destructive actions still require the appropriate level of human sign-off. Peterson's formulation is that anything the company isn't comfortable having an agent handle gets routed to human judgment.

That is the right design, and it is also where the hard questions live. The source does not disclose how Sophos decides which cases are agent-eligible, who calibrates those boundaries over time, or what happens when an agent's recommended action is wrong in a way an analyst doesn't catch. The customer story says boundaries are "calibrated by Sophos analysts" without describing the mechanism. For buyers, the contractual question — who is liable when an agent operating in Authorise mode isolates the wrong host during a business-critical window — is not addressed.

What the announcement does not say

Several things an experienced security buyer would want are absent. There is no disclosure of accuracy metrics: no false positive rate, no false negative rate, no measure of whether agent-produced investigation summaries hold up against analyst re-review. Speed without a quality baseline is only half a result, and in detection and response the cost of a confidently wrong conclusion is asymmetric.

The announcement also does not name which OpenAI models are in use, does not describe what Daybreak actually is as a program — whether it provides early model access, engineering support, custom tuning or some combination — and does not state the time period over which the metrics were gathered or the sample size behind the 38-minute and 89-second averages. Cost is not disclosed, which matters because Sophos's stated rationale is scaling compute instead of scarce headcount. Whether that trade is favorable depends entirely on inference economics that neither party published.

Finally, this is a vendor-published customer story on a vendor's own site. The numbers are Sophos's, relayed by OpenAI, with no independent audit mentioned.

The defender-attacker framing

OpenAI's framing of the story leans on a point worth taking seriously: the same capabilities that speed up defensive triage are diffusing into open-weight models that attackers can run without a provider's usage policy in the way. Peterson's own advice to security leaders is notably unglamorous — focus on fundamentals, patch, and maintain layered defenses including endpoint protection, multifactor authentication and network segmentation — with the caveat that patching only covers vulnerabilities a vendor already knows about. For a story about frontier-model agents, the closing recommendation is essentially 2015-era hygiene, which is a useful corrective to the implied narrative.

For product teams outside security, the transferable lesson is the shape of the deployment rather than the headline number: a narrow, high-volume, well-instrumented workflow; an explicit escalation boundary tied to action reversibility; and an existing human baseline good enough to measure against.

What to watch

• Whether Sophos or any independent party publishes accuracy data — false positives, missed detections, analyst override rates — to sit alongside the speed numbers.
• Whether the 52% end-to-end automation share rises as agent coverage broadens, or plateaus because the remaining cases are structurally harder.
• What OpenAI's Daybreak program actually provides to partners, and whether it becomes a named commercial tier or stays an invite-only arrangement.
• How liability and incident-review terms evolve for customers using Authorise mode, where an agent can act without pre-approval.
• Whether competing MDR providers publish comparable per-case timing metrics, which would turn these numbers into a benchmark rather than a single vendor claim.