Procurement is granting autonomy to security tools whose AI is a single anomaly score

Why it matters

cost The AI premium is paid out of penetration testing, identity hardening, secure development, tabletop exercises and experienced analysts, because the column's premise is that the budget does not grow to fund the label.

exposure A tool that adds ambiguous findings without improving prioritization leaves the exposure with the triage queue, where an intrusion that unfolds slowly is the one most likely to be reached late.

constraint Genuine machine learning bills for data quality, baseline currency and monitoring through mergers and cloud migrations, which forecloses the staffing cut that often justified the purchase.

decision Someone has to be named as the owner of a wrong autonomous action before signature, since the column's position is that the product may be smart without being accountable and the buck stops with management.

The test fits inside one vendor call. Ask which decision the model makes without a human in the loop. Then ask what the observable result is when it makes that decision wrong. The SC Media column supplies the shape of a good answer in its own SOC example: a connection blocked because a rule matched a malicious IP address leaves the analyst a clear next step, while a database query flagged for deviating from an opaque behavioral baseline leaves the analyst holding an alert and none of the reasoning ^[5].

The column is not dismissing probabilistic detection outright; it allows that models surface patterns static rules miss, and locates the failure elsewhere: a tool that cannot explain its reasoning in operational terms produces uncertainty instead of clarity ^[11]. The buying question, then, is whether each output arrives with a reason an analyst can act on at 3am, and with a defined rollback when the action was wrong, not whether the model tests well in a lab.

Now count what the autonomy budget is competing against, using the column's own lists. Most breaches trace back to the same commodity origins: exposed services, stolen credentials, unpatched systems, misconfigured cloud resources, poorly segmented networks, and an employee tricked by a convincing message ^[4]^[12]. Alongside those sit the unglamorous disciplines the column says keep getting starved: asset management, patch governance, privileged access management, backup testing, phishing resistance, incident response planning, and the human side of defense ^[13]^[14]. That is thirteen line items in all ^[15], and none of them needs a model, though all of them keep appearing in incident reports ^[4].

The label itself has a legitimate use, which is worth stating plainly. The same column credits AI where it earns the name, sifting telemetry at machine speed and correlating weak signals across large event volumes ^[16]. What travels with the label is the second-order grant, and that grant is written in headcount and in autonomy scope, not in the datasheet.

Read the source for what it is. This is a Perspectives commentary column with no dataset and no vendor named ^[18], which makes it an argument rather than a measurement. The argument a buyer can act on is the arithmetic: thirteen maintenance obligations and one autonomy purchase draw on the same budget cycle and the same short attention span, and only one of them arrives with a sales engineer.

Claim ledger

Ranked by verification strength, evidence, and original report placement.

The column contrasts two SOC cases: if a system blocks a connection because a rule matched a malicious IP address the workflow remains fairly clear, but if a model flags a database query because it deviates from an opaque behavioral baseline the team may have a harder time deciding what to do next.