MODULE 6 · LESSON 2
Free — no login requiredSign in to track progress, save quiz attempts and enrol in the full course.
Sign in to track progress / enrolDefending With AI, Honestly
Defenders received the same technology as attackers. It is worth being precise about where it pays off, because claims in this market run well ahead of evidence and budgets get spent on the difference.
Where it genuinely helps
Triage at volume. A security team receiving thousands of daily alerts, most of them false, can have AI cluster, rank and summarise them. This is the clearest win available today, because the task is high volume, pattern shaped, and currently performed by expensive people who burn out doing it.
Anomaly detection. Recognising that an account which always signs in from Chennai on weekday mornings is now enumerating the file server at three in the morning from an unfamiliar network. Note that this addresses the Module 3 problem directly: a valid login produces no abnormality, so behavioural patterns are among the few remaining signals.
Speed of comprehension during an incident. Summarising a long log trail, explaining unfamiliar code, or drafting a timeline while an incident is running. Minutes matter, and this is a genuine accelerator.
Finding weaknesses before shipping. Reviewing code for known weak patterns at a scale humans cannot match. Not a replacement for review, but a useful first pass.
Where it will let you down
Confident wrong answers, at the worst moment. An AI that invents a plausible explanation during a live incident can send an entire team down a dead end while the actual intrusion continues. Incident response is exactly where fluent misinformation is most expensive, because decisions are fast and confidence is scarce.
Your security tooling is itself a target. Everything in Module 5 applies here. A detection system that reads attacker supplied content is a system processing hostile input, and attackers know to aim there.
More detections without more capacity. If the constraint is people able to act on alerts, generating more alerts does not help and may hurt. Alert fatigue arrives faster.
It does not substitute for the basics. No detection layer compensates for unpatched systems, shared credentials, or backups nobody has tested. Buying detection while neglecting those is paying for a smoke alarm in a building with no fire exits.
The question to ask a vendor
When someone pitches an AI security product, one question separates the useful from the decorative.
What happens when it is wrong, and how would I know?
A tool that is correct 95 percent of the time and unfalsifiable the other 5 percent is worse than a simpler tool whose reasoning you can check, because you cannot calibrate your trust in it. Ask specifically what a false negative looks like, how false positives surface, and whether an analyst can see why a judgement was made.
Follow it with the Module 2 question: what catches it when this fails? If the answer is that nothing does, you have bought a single point of failure with good marketing.
The honest overall assessment
In the short term, AI has shifted the balance toward attackers. Three reasons, each covered earlier.
Generating a convincing lure is easier than detecting one. The attacker needs to succeed once while the defender must succeed continuously. And the capability AI most degraded, human detection of deception, was load bearing in a great many organisations' defences.
The correct response is not to try to win the detection race. It is to remove the classes of attack that detection was compensating for, which is what every structural recommendation in this course does.
A passkey does not detect a phishing site, it makes the attack impossible. A second approver does not detect a fraudulent instruction, it makes one convinced person insufficient. Least privilege does not detect an intruder, it limits what one is worth. An approval gate does not detect a prompt injection, it caps what a successful one achieves.
That is the through line of the whole course. Where you can convert a detection problem into a structural one, do it, because structure holds while detection erodes.
Both sides gained the same capability. It is fair to ask why that does not simply cancel out.
The tasks are not equally suited to the technology. Generation is a task with many acceptable outputs: any convincing lure will do, and the attacker can generate thousands and keep whichever works. Detection requires a correct judgement about a specific case, with an expensive cost for both error types. Current systems are considerably better at the first kind of task.
The error costs are asymmetric. An attacker whose message is filtered loses nothing and sends another. A defender who wrongly blocks legitimate business creates real cost immediately, which pushes defensive thresholds toward permissiveness. Attackers can tolerate a 99 percent failure rate. Defenders cannot tolerate a 1 percent false positive rate on legitimate mail.
Feedback loops differ in speed. Attackers learn from every attempt, cheaply and immediately: it worked or it did not. Defenders often learn only after a successful breach, which is expensive and rare enough to be a poor training signal.
Deployment speed differs. An attacker adopts a new technique the day it exists. An organisation procures, pilots, integrates and trains, which takes months. The gap between capability existing and capability being deployed defensively is where a lot of incidents happen.
None of this means defence is hopeless, and the pessimistic reading misses the point. It means defence should not be built on winning a detection race that has structural disadvantages. It should be built on controls that do not depend on winning it, which is why passkeys, least privilege, approval gates and dual authorisation appear repeatedly in this course. Those controls are unaffected by every asymmetry listed above, because they do not require anyone, human or machine, to correctly identify an attack.
A vendor demonstrates an AI security tool that correctly identified every attack in their demonstration. What is the most useful question to ask next?
Alert triage
Click to flipClustering and ranking high volume security alerts. The clearest current win for AI on the defensive side.
Click to flip backAI genuinely helps defenders with alert triage, behavioural anomaly detection, and speed of comprehension during incidents. It will let you down through confident wrong answers at the worst moment, by being a target itself, and by producing more alerts than anyone can action. Ask any vendor what happens when the tool is wrong and how you would know. Above all, recognise that the detection race starts with structural disadvantages, so the productive response is to remove the attacks that detection was compensating for. Structure holds while detection erodes.