profile

CybersecurityOS 🛡️

The rule worked. The detection still failed.


Twenty-two seconds. That's the median time, per Mandiant's M-Trends 2026, for initial access to be handed off from one threat actor to the next. One crew gets in. Another takes over.

Now look at your queue. An alert fires. It waits. Someone pivots across four consoles, decides it's probably the backup job again, and closes it. The rule worked. The detection still failed.

For a decade, the constraint in detection engineering was signal creation: writing enough good rules, fast enough. That problem hasn't gone away, but it's no longer the binding one. The bottleneck moved to signal consumption: who or what investigates the alert, and how fast. AI agents are entering that layer, and it changes the job.


Where the field stands

The SANS State of Detection Engineering 2026 (307 practitioners) found 80% of teams barely keeping pace or falling behind. Only 18% are ahead. Cloud-native is the top coverage gap, named 2.5x more than any other environment. Software engineering proficiency is the biggest skills gap. And Gartner has been reported as expecting 75% of SOCs to deploy AI analysts by the end of 2026.

A large share of your alerts will soon be read first by something that isn't a person.

Detections now have two audiences

A detection written for a human says: "User granted high-risk permissions to an unverified third-party app." A good analyst knows what to check next. An agent doesn't. It has what you give it, and a thin detection gets filled in with confident guesses.

Written for an agent, that same OAuth consent-phishing detection carries its own reasoning: context, evidence to collect, expected benign explanations, an ATT&CK mapping (T1528), confidence conditions written as logic, and response guidance including when to hand off to a human.

My position: write the detection for the agent first. One complete enough for an agent is better for humans too. The reverse isn't true.

Glass box, not black box

You can't tune automation you can't inspect. When a black-box system closes an alert, you get a verdict. If it's wrong, you can't tell whether the rule, the enrichment, or the model failed.

It's a governance problem too. "The model decided" won't survive a regulator or a post-incident review. If you can't reconstruct why an alert was closed, it shouldn't be closed automatically.

Detection-as-Code stalls on people, not tools

Most teams agree with Detection-as-Code. Far fewer run it. If a team has never written a unit test or reviewed a pull request, a repo just becomes a place where rules get dumped. A tool purchase is not a skills plan.

Detection moves toward the data

Not everything has to go through the SIEM anymore. Detection is moving into pipelines and data lakes, closer to where telemetry is generated. That cuts premium ingest costs and is one of the more realistic ways to close the cloud-native coverage gap. The catch: rules spread across more places, and without Detection-as-Code discipline, nobody knows what's deployed where.

The new target: your own agents

The AI agents in your SOC are privileged non-human identities. They read sensitive logs. Many can disable accounts, revoke tokens, and close tickets. An attacker who can poison an agent's inputs may be able to steer its decisions. An agent that can close alerts can close the alert about itself. Baseline what each agent does and alert on deviation.

Stop reporting coverage percentage

Coverage measures whether a rule exists, not whether anyone acts on it correctly. Try these instead:

  • Time from intel to deployed detection
  • Percentage of alerts driving a correct action
  • False-positive cost in analyst hours

Five predictions for 2027

  1. A widely used open detection rule format adds fields for agent investigation context.
  2. A publicly disclosed incident involves an attacker manipulating a SOC AI agent to suppress or close alerts.
  3. Cloud-native is still the top coverage gap in the next SANS survey.
  4. A financial services regulator publishes guidance on explainability of AI-driven alert triage.
  5. Detection engineer job postings ask for software engineering skills more often than any single SIEM query language.

Monday morning

For practitioners: write expected benign explanations into your five noisiest rules, rewrite one detection end to end for an agent, and put one rule through real code review with a real test.

For leaders: replace coverage percentage with one of the metrics above, require an evidence trail for every automated closure, and inventory every AI agent with access to your security tooling.


One question for you. If an agent closed an alert in your SOC yesterday, could you explain to your regulator today exactly why? Hit reply. I read every answer.

The full write-up goes deeper on each section: read it on the blog.

​Michael Tayo - CEO/Founder, CyberShield

CybersecurityOS 🛡️

Engineering security from first principles. I'm d0uble 3L, I write CybersecurityOS, where I break down secure-by-design architecture, DevSecOps, cloud security, and emerging-tech risk into practical frameworks for engineers, leaders, and teams. Weekly perspectives, clarity over complexity.

Share this page