Anitha calls it the year of the shortest job description she ever received. “Read the queue,” her manager said, when she joined a mid-size Bengaluru fintech’s security team. The queue was phishing alerts, and it never ended. Two years later her title changed to something that did not exist when she started: agent governor — the person responsible for deciding what the automation on the team may and may not do. Her route from queue reader to that role is a usable map for anyone watching AIOps for Security reshape security careers, because every step on it was a learnable skill rather than a lucky break.
- Skill one: labelling evaluation data blind to the machine’s verdicts.
- Skill two: bound arithmetic that settles arguments in meetings.
- Skill three: turning trust into written, signed gates.
- Skill four: keeping watch after go-live with canaries and drift monitors.
- Skill five: governance fluency — reading frameworks so vendors cannot out-jargon you.
A Queue Reader in Bengaluru
The job at first was volume. Hundreds of alerts a day, each one a small verdict to make: genuine or noise. Anitha became fast, then became suspicious of her own speed, because the days blurred and the mistakes stayed invisible. Nobody audits a human queue reader. That last detail deserves more weight than it appears to carry — it is the exact gap that machine triage now promises to fill, and the exact gap this whole field keeps forgetting to measure when it does.
The Meeting Where AIOps for Security Arrived
A vendor demonstration landed in her inbox in the usual way: a deck promising that an AI agent could close most of the team’s phishing alerts automatically, paying for itself in saved hours. Her manager was enthusiastic. Anitha’s instinct was the one every careful buyer in this market needs — she asked how the tool had been measured, on whose mail, and by whom. The answers were vague, which turned out to be the most useful discovery of her quarter. She went looking for how serious teams evaluate such claims, found the vocabulary of the emerging field — the Ops Atlas material describing AIOps for Security and its evidence gates was her reference throughout — and came back to the meeting with a proposal that made her indispensable: let the agent work, but let nobody trust it until it passes a test we design.
First Skill: Labelling Without Looking
Her first task was building the test, and it began with labels. She and a colleague graded several hundred recent alerts as genuine or noise, working independently, without seeing what the vendor’s model claimed about any of them. Blind labelling sounds simple and is strangely rare; evaluations keep getting graded by whoever built the system being graded, and research stretching from 2019 through 2022 — the TESSERACT findings — showed how spatial and temporal bias in test data could float an F1 score to a deceitful 0.99. Anitha’s rule was blunter: if the labels and the verdicts could see each other, the evaluation does not exist.
Second Skill: Arithmetic That Ends Arguments
The second skill was small-station maths. Anitha learned that a spotless record proves less than it advertises: zero errors across n cases leaves the true error rate bounded near three-over-n. That is why a clean demo convinces statisticians of nothing. She learned the sample demands by heart — roughly 236 graded cases before anyone may claim a 2% miss ceiling once a single miss has appeared, and about 313 once two misses show up. And she learned the trap of spot-checks: audit a random twentieth of a quarter’s closes — around ninety items out of some 1,845 — find nothing wrong, and a miss rate near 38% remains entirely possible. In meetings, whoever holds the arithmetic holds the decision. She made sure it was her.
Third Skill: Writing Gates Down
The third skill was turning trust into paperwork. Anitha drafted a one-page policy that granted autonomy rung by rung, per action class rather than per product. Watching came free. Recommending required the labelled pile her team had built — 200-plus graded alerts holding 30-plus genuine cases. Auto-closing required a demonstrated miss bound no worse than 2%, proven on 150 or more true positives, plus a blind holdout, canaries, and a one-click reopen. Any reversible action required a false-action ceiling near 1%, proven over 300-plus verdicts on items the system would have acted upon, plus a rollback that had actually been rehearsed. Irreversible actions stayed permanently human. When the vendor asked which of these were negotiable, she enjoyed saying: none.
Fourth Skill: Keeping Watch After Go-Live
The fourth skill was unglamorous and decisive: post-deployment vigilance. The policy planted roughly fifty canary alerts a quarter — messages crafted as genuine threats, slipped into the flow where nobody could cherry-pick around them — under a three-strike rule: if the agent closed a third canary, its auto-close rights ended that day. Drift monitors watched the inputs that move most, because a model calibrated on March mail ages badly by September. And Anitha reviewed category-level numbers monthly rather than blended ones, having read how a pooled figure once hid 42.3% of cloud-identity misses behind a pleasant 9.1% phishing result. None of this required genius. All of it required showing up.
The Fifth Skill Nobody Puts on Job Ads
Reading governance documents turned out to be the fifth skill, and the least expected one. The NIST AI Risk Management Framework — first issued as version 1.0 in 2023 and now mid-revision — gave her a vocabulary her directors already recognised; the Generative AI Profile called AI 600-1 came in 2024, with the E2025 revision of AI 100-2 following in March 2025. The NCSC’s papers — Deploying AI Systems Securely from April 2024, AI Data Security from May 2025 — turned vague unease about data routes into specific questions: where do prompts go, is sharing on by default, what leaves the tenant. One line she quotes to every new joiner: Security Copilot ships with sharing of customer data switched on unless the organisation holds M365 E5 or E7 licensing.
What AIOps for Security Careers Actually Reward
Eighteen months in, the queue is quieter and Anitha’s judgement is the scarcest resource on the team — the agent proposes, she decides what it may do, and the decision scorecard she keeps rates readiness from 0 to 3 across weighted dimensions, max 27: above 14 means shadow mode, 9 to 14 means a time-boxed pilot, below 9 means no. Careers in this field reward exactly that stack: honest labels, bound arithmetic, written gates, live canaries, and fluent governance. They do not reward prompt-trick memorisation or vendor loyalty, both of which age in months.
The balance matters too, and this being Bangalore, she preaches it with the city’s own calendar: close the laptop after a canary review and get out — the picks in Bangalore music events or the round-up of theatre experiences around Bangalore are her standing recommendations for decompressing after audit weeks. For the technical ladder itself, the team onboards new analysts with a step-by-step DevOps career skills guide, since governing agents starts with understanding the systems they watch.
Anitha’s summary fits on a badge: the machines got faster at triage; humans did not become optional. Somebody still has to decide what they may close, and that somebody — the governor — now has the best seat in the house.
