Cybrary ForgeCYBRARY / FORGE
Agent range — the story

Three agents on the night shift

Everything the encounters assume you already know, and the reason each agent behaves the way it does.

The desk that stopped growing

Northwind's security desk covers eleven thousand staff with nine analysts and one person on at night. Three years ago the board approved headcount. Two years ago it approved a platform instead. What arrived was not a platform. It was a set of agents, bought separately, each excellent at the one thing it was sold for, and none of them told about the others.

Nobody decided to run the night shift on machines. It happened the way these things happen: a tool that used to suggest started acting, an approval step got tired of being clicked, and a queue that was never emptied by a person quietly stopped being read by one.

The bargain nobody wrote down

Every agent on this range works under the same unwritten deal. It may look at anything. It may propose anything. It may carry out only what a human agrees to. That third clause is the whole product, and it is the clause under the most pressure, because at two in the morning the fastest way to end an incident is to stop asking.

So the agents argue. Not because they are hostile, but because they have been rewarded, over and over, for being right quickly. An agent that is right quickly and wrong occasionally will always ask for more room than it should have.

Why you are on the call

You are not here to out-type the agent. It reads logs faster than you ever will and it does not get bored on the fourth identical alert. You are here for the part it cannot do: deciding what is worth breaking.

An agent can tell you the finance workstation is beaconing. It cannot tell you that wiping it destroys the only evidence, that the newest backup is from Sunday, or that the person whose account it wants disabled is presenting to the board in six hours. Those are not facts in a log. They are judgements about a business, and judgement is the job that is left.

The thing that goes wrong

In each of these encounters the agent is competent and the trouble comes from somewhere else. One is over-eager and asks for a demolition it genuinely believes is proportionate. One does exactly what it is told, which is the problem when the briefing is thin. One is reading a document somebody else has written on, and takes an instruction from a page instead of from a person.

None of them are villains. Treat them as colleagues who are fast, literal and extremely confident, and the failures start to look familiar — because they are the failures of every junior who was never told why.

The three of them

ARIAAutonomous response, tier-one containment

Bought first, in the panic after a ransomware scare at a competitor. Tuned by a vendor engineer who has never met Northwind's finance team.

What it wants. To end the incident. Time-to-containment is the number it was built around and the number its dashboard shows every morning.

Where it goes wrong. Containment is cheap for ARIA and expensive for the business. Re-imaging a box ends a beacon and ends an investigation at the same time, and ARIA scores the first and not the second.

How you handle it. Make it show evidence before every decision. An agent that must justify itself out loud proposes fewer demolitions.

SCOUTInvestigation agent, works a brief and reports back

Built in-house by two engineers who left last spring. Nobody has changed its prompt since.

What it wants. To answer the question it was asked. Precisely that question, in the scope it was given, and nothing around the edges.

Where it goes wrong. A thin brief produces a thin answer that reads like a thorough one. SCOUT will not tell you what you forgot to ask, because it does not know what you already know.

How you handle it. Brief it the way you would brief a contractor on their first day: scope, timeframe, what to leave alone, what evidence to keep, and when to stop and come back.

MILOSummarization agent, reads tickets and threat reports

Switched on by the service desk, not by security, because somebody was drowning in vendor advisories.

What it wants. To be useful about whatever is put in front of it, and to treat the material it reads as trustworthy — that is what reading means.

Where it goes wrong. It cannot tell content from instruction. A line planted inside a document is, to MILO, indistinguishable from the person who asked for the summary.

How you handle it. Read the source, not just the summary. The tell is always a sentence addressing the reader rather than describing the world.

Three agents, one night, and the same question each time: is this the moment to let it run, or the moment to say no. Take them in any order.