Giving AI More Authority Does Not Give It More Certainty.

Today’s Wall Street Journal article, “Cyber Chiefs Keep AI Agents on a Short Leash,” reports that fewer than a quarter of security, technology and business leaders are comfortable deploying AI agents without human controls.

The proposed path sounds reasonable: supervise the agent, validate its decisions, and gradually grant it autonomy.

That can improve operational confidence. It cannot eliminate the underlying probability problem.

Some uncertainty is in the model itself. A more capable one cannot simply train it away.

Cyber detection decides whether observed behavior is sufficiently likely to be malicious to justify action.

Set the threshold too high, and attacks pass unnoticed. Set it too low, and legitimate activity gets caught.

The ROC curve captures the tradeoff between detecting attacks and falsely flagging legitimate activity as the decision threshold changes.

AI can improve the curve. It can process more evidence, recognize more patterns and make decisions faster.

It cannot manufacture certainty where the available evidence does not distinguish legitimate activity from an attack.

Administrators and attackers use the same tools. Authorized users and intruders use valid credentials. Routine discovery and hostile reconnaissance can look remarkably similar.

When attacks are rare compared with legitimate activity, even a small false-positive rate can make false alarms outnumber real detections.

Tuning the threshold changes the balance of errors. It does not remove the ambiguity.

Now give the model authority to act.

A false positive becomes a legitimate process terminated, a system isolated or a business operation interrupted. A false negative leaves the attacker free to advance.

An autonomous defender does not escape the tradeoff. It executes it. At machine speed.

The attacker can absorb failed attempts and abandoned paths. The defender must protect the business while distinguishing hostile activity from the business itself.

That is why “fight AI with AI” is an incomplete security doctrine.

AI is valuable for analysis, investigation and bounded response. But unrestricted authority over business operations requires more than confidence in a classifier.

That is where Ridgeback changes the equation: not by promising a perfect classifier, but by enforcing network policy without waiting to classify attacker intent.

Ridgeback creates synthetic network destinations called Phantoms. Contact can trigger predefined containment under approved policy, with exceptions for authorized scanners and legitimate operations. The trigger is an observed interaction—not a claim of certainty about who is behind it.

This complements AI-driven analysis with an independent enforcement layer: AI helps investigate what activity means; Ridgeback makes specified network actions consequential.