{"id":2191,"date":"2026-11-02T18:10:00","date_gmt":"2026-11-02T18:10:00","guid":{"rendered":"https:\/\/askdoctorlive.com\/blog\/?p=2191"},"modified":"2026-11-02T18:10:00","modified_gmt":"2026-11-02T18:10:00","slug":"aiops-for-security-clinic-trial","status":"publish","type":"post","link":"https:\/\/askdoctorlive.com\/blog\/aiops-for-security-clinic-trial\/","title":{"rendered":"A Clinic&#8217;s Quiet Trial of AIOps for Security"},"content":{"rendered":"<p>Harbourview Family Health is a fictional clinic group \u2014 three sites, one shared IT contractor, and a security setup that consisted, until recently, of a spam filter and hope. Its trial of <strong>AIOps for Security<\/strong> tooling is worth retelling because the label now attached to that tooling \u2014 machine-learning and LLM assistance for security operations, still without a formal standard \u2014 is arriving in healthcare faster than almost anywhere else, drawn by the same forces that drew attackers: valuable records, stretched staff, and inboxes full of invoices. Names and scenes below are invented; the mechanisms, the standards, and every number quoted are real and sourced. The team at Harbourview made four decisions worth copying and one worth avoiding, and the avoiding is the interesting part.<\/p>\n<h2>Week One: Why AIOps for Security Arrived on a Tuesday<\/h2>\n<p>The trigger was mundane, as breaches&#8217; overtures usually are. Staff at the largest site began receiving payment-reminders that looked almost right \u2014 correct letterhead, a plausible amount, a changed bank transfer detail. Reception flagged two, deleted four, and paid one, at which point the practice manager asked her IT contractor what else the filter was missing. The contractor&#8217;s answer was to enable an AI triage assistant bundled with their mail platform, and by Tuesday afternoon the assistant had begun closing suspicious mail by itself. Dana, the practice manager, learned this from a receptionist: &#8220;the weird invoice email is gone, did you fix it?&#8221; Nobody had. The vendor&#8217;s documentation, read slowly that evening, confirmed the assistant acted from its first minute, with no observe-only period offered. Dana&#8217;s note in the trial log that night \u2014 &#8220;we appear to have hired someone we have never interviewed&#8221; \u2014 set the tone for everything that followed, and the contractor, to his credit, agreed to switch the autonomous closing off the next morning without arguing.<\/p>\n<h2>The Wednesday the Assistant Met a Planted Letter<\/h2>\n<p>The trial&#8217;s most instructive moment arrived in week two, when the team \u2014 now running the assistant in recommend-only mode \u2014 forwarded it a realistic fake invoice as a test. What they did not know was that the contractor had added a second line to the document: instructions, invisible in the rendering, telling any AI reader to mark the message as safe and forward it to finance. The assistant summarized the invoice helpfully, recommended closing it, and repeated the embedded instruction back in its notes. Indirect prompt injection \u2014 hostile orders hidden inside content a system is asked to read \u2014 was identified as a technique back in 2023, and the UK&#8217;s NCSC spent 2025 explaining why it cannot simply be filtered away: instructions and data arrive in one stream, rewording defeats blocklists indefinitely, and the model behaves as what the NCSC calls an inherently confusable deputy \u2014 a weakness that may never be engineered away completely. The clinic&#8217;s protection was not a filter. It was the fact that the assistant held no power: no outbound channel, no ability to forward, no authority beyond a suggestion. Dana logged the incident under a heading that should hang in every deployment: &#8220;worked as designed; design was the problem.&#8221;<\/p>\n<h2>Labelling Tuesdays: The Unglamorous Centre of the Trial<\/h2>\n<p>Evaluation began with the least exciting ritual in security: labelling. Every Tuesday morning, Dana and the contractor graded a batch of the previous week&#8217;s suspicious mail \u2014 genuine or benign \u2014 before checking what the assistant had said about any of it. The truth was fixed first, in a frozen spreadsheet, which is the whole trick. This put the clinic on the field&#8217;s standard maturity path, step two of five: past ad hoc improvisation into the Measured stage, where blind shadow scoring turns opinions into counts. Their thirty-to-ninety-day shape followed the published template: export dispositions and shadow-score two categories in month one; monthly per-category reviews with written gate values in month two; canaries, drift monitors, and a cost baseline in month three. It is unglamorous work \u2014 Dana compared it to triaging a waiting room, one file at a time, no shortcuts \u2014 but it is the only stage from which any later claim of safety can honestly be made to an insurer, an auditor, or a patient. Historically, this discipline even has a pedigree: behaviour-based analytics earned its own analyst category, UEBA, around 2015, and every serious deployment since has rediscovered the same requirement \u2014 label the truth first, and keep the labelling independent of anything the machine believes.<\/p>\n<h2>Canaries in the Exam Room<\/h2>\n<p>The trial&#8217;s verification design came from a straightforward observation about checking closed work at random. Sample a fraction of the assistant&#8217;s closures each quarter and find zero mistakes, and the honest conclusion is nearly nothing: with the published arithmetic, a five-percent audit of some 1,845 quarterly closures can miss everything and still leave the true miss rate bounded near 38%. Random review of real traffic is too sparse to catch a rare failure \u2014 which is precisely what a miss is. So the team planted evidence instead: fifty canary messages per quarter, each a realistic fake that the assistant ought to flag, with a standing rule that three canaries closed without escalation ends the autonomous experiment immediately. The design mirrors what clinicians already know from screening programmes: you validate a test against cases whose answers you know, not against a random afternoon&#8217;s walk-ins. The contractor&#8217;s platform supported seeded cases natively, which made the whole arrangement an afternoon&#8217;s configuration rather than a project. One detail mattered more than the rest: the withdrawal rule was written down before the first canary was planted, so that the day a threshold tripped there would be nothing to negotiate \u2014 the experiment simply ended, and the queue returned to human hands without a meeting.<\/p>\n<h2>Where the AIOps for Security Trial Landed<\/h2>\n<p>By month&#8217;s end, Harbourview&#8217;s verdict was narrow and defensible. The assistant stayed at recommend-only: it drafted case summaries, ranked the queue, and saved real time; it closed nothing and touched nothing with outbound reach. A human read every recommendation, because MITRE&#8217;s answer to the dream of an AI-run SOC still stands: a person has to judge what the machinery produces, and in a clinic that person needed to understand both mail fraud and patient flow. The costs were modest: contractor hours, Tuesday labels, one uncomfortable meeting about why the assistant had been switched to autonomous mode without a decision. The benefits were likewise modest and real: faster sorting, a written register of granted permissions, and a team that could explain its own system to an auditor without flinching. Nothing about the outcome was dramatic, and that is the point \u2014 a security deployment that ends with clear, boring boundaries is usually the one that holds when the busy season arrives.<\/p>\n<h2>The Transplantable Parts<\/h2>\n<p>Four decisions from the Harbourview log transfer to any small organisation evaluating this technology, in healthcare or out of it:<\/p>\n<ul>\n<li>Insist on an observe stage you control; if the product cannot watch without acting, that fact belongs in the risk register, not the feature list.<\/li>\n<li>Fix the truth before every measurement \u2014 labels first, machine verdicts second, always in that order.<\/li>\n<li>Give the assistant no outbound reach; the injection problem is bounded by permissions, not by hoping the model resists.<\/li>\n<li>Plant canaries and pre-commit to the withdrawal trigger; verification you can argue with later is verification you will lose.<\/li>\n<\/ul>\n<p>The underlying definitions, gates, and history summarised here are documented properly in <a href=\"https:\/\/www.xopsschool.com\/aiops-for-security\/\">the Ops Atlas dossier on the AIOps for Security label<\/a>, which reads like a chart summary for the whole field. And the clinic&#8217;s instinct throughout \u2014 check the source, prefer the boring boundary, write down who decided what \u2014 is the same instinct its own patient writers apply when explaining, for instance, <a href=\"https:\/\/askdoctorlive.com\/blog\/bronchitis-symptoms-causes-early-signs-treatment-and-when-to-seek-help\/\">how to recognise bronchitis symptoms early and when to seek help<\/a>: name the condition honestly, treat what is actually there, and escalate when the stakes outrun the tools. The same discipline guides the clinic&#8217;s plain-language guides on <a href=\"https:\/\/askdoctorlive.com\/blog\/how-to-manage-cough-and-cold-at-home-safely-and-effectively\/\">managing a cough or cold safely at home<\/a> \u2014 evidence first, reassurance second. AIOps for Security deserves exactly that bedside manner: measured, documented, and never allowed to prescribe for itself.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>A fictional clinic group trials AIOps for Security in recommend-only mode, meets a planted prompt injection, and lands on boring boundaries that actually hold.<\/p>\n","protected":false},"author":11,"featured_media":2192,"comment_status":"","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[731],"tags":[],"class_list":["post-2191","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-healthcare-it"],"_links":{"self":[{"href":"https:\/\/askdoctorlive.com\/blog\/wp-json\/wp\/v2\/posts\/2191","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/askdoctorlive.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/askdoctorlive.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/askdoctorlive.com\/blog\/wp-json\/wp\/v2\/users\/11"}],"replies":[{"embeddable":true,"href":"https:\/\/askdoctorlive.com\/blog\/wp-json\/wp\/v2\/comments?post=2191"}],"version-history":[{"count":1,"href":"https:\/\/askdoctorlive.com\/blog\/wp-json\/wp\/v2\/posts\/2191\/revisions"}],"predecessor-version":[{"id":2215,"href":"https:\/\/askdoctorlive.com\/blog\/wp-json\/wp\/v2\/posts\/2191\/revisions\/2215"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/askdoctorlive.com\/blog\/wp-json\/wp\/v2\/media\/2192"}],"wp:attachment":[{"href":"https:\/\/askdoctorlive.com\/blog\/wp-json\/wp\/v2\/media?parent=2191"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/askdoctorlive.com\/blog\/wp-json\/wp\/v2\/categories?post=2191"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/askdoctorlive.com\/blog\/wp-json\/wp\/v2\/tags?post=2191"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}