Security Teams Hit Human Limits As Alert Volumes Force Move To Automated Containment Frameworks

The idea of alert fatigue didn't start with AI, but attackers using machine learning aren't making it any easier for security teams. Attackers now routinely operationalize AI to build stealth directly into their playbooks, launching campaigns at unprecedented scale and speed. With automated techniques becoming a standard part of the attacker toolkit, human-in-the-loop detection cannot separate signal from noise fast enough to stop modern malware. But for many organizations, the real friction lies outside the technology: in undefined policies, undocumented tribal knowledge, and the lack of authority to actually hit the kill switch.

To unpack where the modern security operating model is breaking down, we sat down with Stewart Alpert, an industry expert and CIO/CTO who previously held the dual roles of CTO and CISO at a billion-dollar global hospitality and transportation giant Hornblower Group spanning hundreds of locations and offices worldwide. Alpert was candid about both the technical evolution of threat detection and the organizational dysfunction that prevents most security programs from translating tooling investments into measurable risk reduction.

"We are getting hundreds of thousands to millions of events per day coming into the SIEM," said Alpert. "The human capacity to understand them has long since been exceeded. That is exactly why security has become an AI-speed problem, not a human-scale one."

Keeping Pace with Automated Threats

Machine learning has fundamentally changed how experts look at threat management, according to Alpert. Early machine-learning approaches scored individual events as normal or anomalous; the next generation layered AI-driven pattern recognition across correlated events at speeds faster than humans. That progress has been met by attackers using the same techniques, enabling them to shift their patterns faster than defenders can manually adapt. "It's AI versus AI, and that's kind of where we sit at this point in time," Alpert explained, echoing what security chiefs across industries are already saying.

Waiting for events to land in the SIEM before correlation begins is already too late. Detection has to move closer to the edge, and the human has to come out of the bottleneck so that automation stops the bleeding before damage spreads. That logic is driving the move toward agentic log intelligence and automated security workflows that compress detection and response into a single machine-speed loop. Recent commentary from Mandia, Stamos, and Adamski at RSAC reinforces the point: organizations that cannot match attacker velocity will absorb damage they did not need to take.

Mind the Response Gap

Even as detection improves, the response side of the kill chain remains stubbornly human. Algorithmic SIEMs and agentic SOC tooling from vendors like Palo Alto, Darktrace, and Google can correlate events at speeds analysts cannot match, and Microsoft's recent work on autonomous defense and expert-led services points the same direction. But once an incident is raised, response is still largely a human endeavor—and that gap is where attackers buy time. Alpert argued automation has to extend past detection and into containment, replacing brittle, ticket-driven workflows with policy-governed actions that execute the moment conditions are met. That is why the conversation around SIEM replacement has moved from theoretical to operational.

Missing the Kill Switch

The harder problem is that most security organizations do not actually have the authority to act, regardless of what their tooling can do. "Most security orgs, and in fact many technology orgs inside of larger businesses today, don't have the authority to shut things down because they're not close enough to the business," Alpert said. He offered a familiar scenario: a customer-service workstation gets compromised and starts behaving badly, but cutting it off may tank inbound revenue and damage the brand, and the CISO rarely owns that call. "How is he going to automate the authority to turn that off?" he asked.

Until someone explicitly owns the kill switch and the dollars-and-cents tradeoffs that come with using it, no amount of agentic automation will close the loop. The fix is organizational before it is technical: clarify decision rights, document the policies governing automated action, and build the agentic log management and risk-quantification practices that enable those decisions to execute at machine speed. The FAIR framework is one of the cleanest ways Alpert has seen to translate exposure into the financial language boards and CFOs actually act on.

Confidently Incorrect

Alpert was blunt about the limits of trusting AI output without guardrails. He recounted a recent session in which a generic LLM gave him an answer he knew to be wrong; when he asked for its confidence, it returned 0.6, and he realized he had skipped his usual default prompt, which instructed the tool not to surface anything below a higher threshold.

The lesson generalizes directly to security automation: AI produces results that match the rigor of the requirements it is given, and nothing more. "If you are running with default training, default settings, you can expect that you're going to get very mediocre results," he warned. Buying CrowdStrike, Darktrace, or any agentic SOC platform and assuming it will work out of the box is, in his mind, the same as hiring a person sight unseen and assuming they will know your business.

The Bus Factor and the Appliance Trap

Alpert reserved his sharpest critique for "appliance people," such as CIOs, CTOs, and CISOs, who buy one of every Gartner-recommended acronym without ever measuring ROI or tying the investment to a specific business risk. "Oh yeah, 'Gartner said I need a SIEM,' but they've never sat down and measured what the ROI is on it," he said.

The right starting point is the inverse: a proper risk analysis, ideally run both internally and by an outside party to neutralize bias, mapped against the assets that carry the most business value, with a Venn diagram of existing coverage overlaid on top. From there, leaders can quantify in dollars and cents what each gap is actually worth closing.

Without that grounding in business risk, governed automation, and clear decision-making authority, Alpert concluded that AI does not save security teams from their old mistakes. The organizations 'winning' in 2026 will be the ones that define risk in dollars and cents, train and govern AI to align with those priorities, and place automation where it can act earliest and most decisively. "It really comes down to the fallacy of 'I think I should do this,' and instead what I really should do is invest the money where the business has the most to lose. We're just making the same mistakes faster."