Have you ever wondered what would happen if one day, the security software in your company decided that a dangerous alert was “benign” and closed it without human intervention? It sounds like a distant scenario, but that is precisely what is happening right now.
In less than a month, an entirely new class of AI models went from “non-existent” to “everywhere.” TypeSafe AI launched Jev in mid-September, calling it a “System One” model inspired by Kahneman’s fast thinking and intuition. OpenAI announced the Decisions API at DevDay and opened public beta in early October. Cribl released cribl-decision, an open LoRA adapter that can run locally. And names like Laya have begun appearing in open-source tools.
The catch: these models do not write essays. You feed them a state and a set of strongly typed questions. They return a choice, a score, or a yes/no answer with a probability. For Security Operations Center (SOC) operations, this could be one of the best pieces of news of the year. But it could also be one of the most dangerous traps.
Why the SOC Should Be Excited About AI Decision Models
Let’s start with the legitimate reasons for optimism. Three characteristics make decision models a natural fit for alert triage in the SOC.
1. You Have a Number to Automate
A strongly typed answer with a calibrated probability is something you can set thresholds on. Ask the model if an alert is benign, and you get back a probability. Auto-close if above 0.97. Below 0.6, treat it as a likely incident and escalate. Anything in between goes to an analyst for closer review. TypeSafe even suggests tighter thresholds for irreversible actions — a clear contrast between “disabling an account” and “adding a comment to a ticket.”
2. They Are Cheap Enough for Every Alert
Jev is priced at $42 per billion input tokens, with responses in 70 to 500 milliseconds. Suppose an enterprise has 450,000 alerts per year and you feed a few thousand tokens into each alert. The model bill comes out to around one hundred dollars a year. The economic argument for “sampling” and reviewing only high-severity alerts just vanishes.
3. Every Verdict Becomes a Log Record
The answer space is fixed beforehand, and the output carries its own probability. You can record every decision, replay it, and compare it with what actually happened. That is much easier to audit than a paragraph of reasoning generated by AI.
In theory, this is a massive leap forward. But theory and reality rarely walk the same path.
The Deadly Trap: When the Input Is Wrong, Everything Is Wrong
This is the part where I, having spent 20 years witnessing how many times “perfect technology” fails in practice, want you to pay special attention.
When vendors say these models “can’t hallucinate,” they mean the output always matches the schema you defined. A correctly formatted answer can still be wrong in substance. And a calibrated guess is still just a guess.
The old rule still holds: Garbage in, garbage out. Calibration is a promise about a distribution. When a well-calibrated model says 90%, it means that on inputs looking like this, it is right about 90% of the time. So the critical question is: what does “like this” look like?
If the input is just an alert title and a severity field, the model is calibrated on alerts with that title and severity. It knows nothing about the host.
And here is why that matters in practice.
Data from a 2026 AI SOC report, covering 25 million alerts triaged across customer bases in 2025, reveals a terrifying figure. Out of those, 82,000 alerts went through a direct live memory scan on the endpoint. Roughly 2,600 of those machines had active infections. And 51% of the infected machines had been reported by the EDR as remediated.
Think about what a decision model sees for one of those alerts: “Malware detected, threat isolated, low severity.” A perfectly calibrated model will close it with high confidence because most alerts looking like that are indeed fine. The model did its job correctly. The input was wrong.
The severity issue is similar at a larger scale. Nearly 1% of actual incidents in the data traced back to alerts at the lowest severity level. With an average of 450,000 alerts a year, that is about 54 real threats, roughly once a week. Cribl’s model openly admits it: it can be overconfident and sensitive to prompt phrasing and choices.
The Less-Discussed Risks: When AI Becomes an “Insider”
The story does not stop at the input problem, however. There is a deeper layer of risk that few want to face.
Over the past 18 months, the SOC analyst assistant has become the highest-privileged component in the modern security operations stack. It reads every SIEM alert, correlates data, and increasingly makes decisions. An AI assistant whose decision logic is compromised — via prompt injections in the alerts it reads, training data poisoning within the model itself, or misconfigured tool permissions — can become an insider threat more dangerous than any external attacker.
And there is a governance paradox that researchers call the “impossible triangle of governance” in generative AI: there is a persistent tension between maximizing efficiency, protecting the legitimacy of decisions, and controlling risk. You cannot maximize all three at once.
Furthermore, multi-agent decision-making models create their own governance challenges: emergent behavior, tacit coordination, cascading systemic risks, and difficulties in attribution. When a chain of AI agents makes a wrong decision, who is responsible? The model developer? The deployment team? Or the system itself — which has no legal standing?
So What Should Enterprises Do? 5 Survival Principles
Facing a landscape that is as promising as it is frightening, I advise a sober approach. Here are the principles that any organization considering the deployment of AI decision models in the SOC should keep in mind.
1. Never Let the Model Decide for You at Irreversible Points
Let the model suggest, but humans must be the final approver for irreversible actions. A recent study showed that only 6% of professionals report using AI to automate decision-making with very little human oversight. The majority still view AI as an assistive tool, not a replacement. That is a ratio worth learning from.
2. Calibration Is Only as Good as the Input
Never forget the lesson of the 51% infected machines that the EDR reported as remediated. Ensure your input data — not just titles and severity, but also host context, history, and behavior — is enriched before feeding it to the model. A perfect model with bad input still makes bad decisions.
3. Build Runtime Governance Layers, Not Just Pre- and Post-Deployment
Current governance frameworks mostly regulate AI before deployment and after an incident occurs, leaving the execution window in between unaddressed. Build real-time intervention mechanisms: automated thresholds, emergency kill switches, and the ability for humans to take over at any moment.
4. Log Everything, But Don’t Just Store It
The auditability of decision models is a huge advantage. But if you do not actually review the decision logs, you are only storing data, not governing. Establish regular review processes for model decisions, especially those with high impact.
5. Prepare for Multi-Agent Complexity
When a single decision model is already hard to manage, imagine a system of multiple models interacting with one another. Emergent behavior and cascading risks are challenges that current governance frameworks are not ready to handle. Start building governance capabilities for that future right now.
Why This Matters for Vietnamese Businesses
You might think: “These models just launched in the US, do they have anything to do with Vietnam?”
The answer is: yes, and sooner than you think.
Modern SOC platforms — whether from international or domestic providers — are all integrating AI into their alert triage workflows. When you use a SOC service, chances are a portion of that workflow is already automated by AI. As decision models become cheaper and more common, they will rapidly find their way into solutions for small and medium-sized enterprises in Vietnam.
This brings both opportunities and risks. The opportunity is that Vietnamese businesses can access advanced cybersecurity analysis capabilities at a lower cost than ever before. The risk is that without understanding the model’s limitations, businesses may make cybersecurity decisions based on guesswork packaged in an appearance of precision.
In a market where many enterprises still lack deep cybersecurity personnel, reliance on AI can become a double-edged sword. AI can help fill the gap, but it can also create a false sense of security if not properly supervised.
When “Accurate” Does Not Mean “Right”
I have watched the cybersecurity industry go through many revolutions: from firewalls to IDS/IPS, from SIEM to EDR, from automation to AI. Each wave promised to solve the problems of the previous wave, and each wave created new problems.
AI decision models are a genuine step forward. They solve a very real problem: alert fatigue, analysis costs, and the need to respond faster. But like every powerful tool, they are not immune to the fundamental rule of computer science: garbage in, garbage out.
And there is a truth I want you to remember by heart: a perfectly calibrated model can still be wrong. It does not tell you it is wrong. It tells you it is confident. In cybersecurity, blind confidence can be your most dangerous enemy.
Use these models. Leverage their cost and speed advantages. But never forget that behind every probability number is a set of assumptions about the world — and the real world is always more complex than any training dataset.
If your company deploys an AI decision model tomorrow to automatically close cybersecurity alerts, where will you set your confidence threshold — and what will you do when the model says “safe” to an alert that your intuition tells you something is wrong with? Share your perspective below, because in an era where machines make more and more decisions, the ability to ask the right questions at the right time may be the ultimate human survival skill!
0 Comment
Add your comment to this article