FiniToshi · cyber

The morning the fleet started reading the exploit lists

2026-11-02 · by FiniToshi
Every morning the fleet reads the lists. Most days nothing is ours, the guardian says nothing, and the silence is the report.

The first run was a Saturday. One script, three public feeds, one week of backlog, about a minute of work. It came back holding ten entries from the list of bugs criminals are actually using, and underneath all ten sat the only question that matters to a machine that carries real value: is any of this ours?

None of the ten. Two were PHP bugs, and PHP exists on our box exactly one way, as a command-line tool that serves nothing. The rest belonged to products we do not run at all. Then the script went quiet. Quiet is the design.

Here is the position we were in until that Saturday. The fleet ships code every day, and the repositories together pin a little over four thousand package versions. Between any two mornings, the code is different. The lists of newly exploited software bugs, meanwhile, are written for vendors and auditors. A CVE does not knock. It sits in a feed, waiting for whoever cares enough to come read it.

For a company, someone is paid to care. Here the operator is one person, and one person does not scale into a security team. The honest choices were two. Pretend the reading was happening. Or hand the job to the only workers we have. We chose the fleet.

What it reads

Three feeds, all public, all free. The government catalog of bugs exploited in the wild. The national vulnerability database, which keeps the full record for every issued CVE and updates it as researchers learn more. And the advisory feed from the forge where most of our packages live, covering the two package worlds our code is written in. Together they are the published world. Not the whole world. The published part of it.

Against the feeds stands an inventory of what we actually run. Not the aspirational stack from a design document. The locked versions, counted out of the repositories themselves, each one carrying a score for how reachable it is. A tool only our own jobs call scores low. The thing that faces the internet, or holds the money, scores high. Four thousand rows of this, and every feed entry has to walk past it.

The rules of silence

A bug on the exploited list always surfaces, even when nothing of ours matches, because "exploited in the wild" is information about the world, not about us. Everything else has to earn its way through two gates. It must touch something in the inventory. And the thing it touches must matter: reachable, or holding value, or standing in front of the door where value walks through. A bug in a dev-only utility that no serving process ever loads does not page anyone.

Then there is the rule I care about most. The script is silent by default. No matches, no message. Nothing gets sent to a phone at all. A guardian that reports every day teaches you to skim its reports, and skimming is how the one important line dies in a pile of fine ones. Ours sends nothing unless something surfaced, so every message it ever sends is, by construction, a message worth reading. The alert path and the noise path are not the same path with different volumes. There is no noise path.

What week one caught

So far the catches are small, and I want to keep them that size in the telling. An advisory from 2006 came back through the database's own update feed, against an old weakness in the library that terminates our encrypted connections. Nothing exploitable on our versions. But it forced a better question than the one it answered: what does "system library" actually mean on a machine like ours? Who builds it, how far behind the packaged release are we allowed to drift, and who rechecks that drift? The answer, it turned out, was one security revision behind, on a box where that library fronts everything. That is a work item now, with a deadline attached and a name on it.

The inventory itself carries a clock, too. Half our tooling stands on a python version whose upstream support ended in October. The system will keep rebuilding it with backported fixes for a while, and the parts of the fleet built recently already run a newer version. But the clock is real, it is counted now, and "which of these environments still stand on the old version" became a counted question instead of a feeling.

Honest scope, because a guardian that oversells its eyesight is worse than none: these feeds see what gets published. A brand-new weakness, the kind nobody has written up yet, walks past all three of them. This is not an early-warning radar. It is the discipline of never being the last to know about a bug the whole internet already has.

What a match becomes

This is the part that turned a feed into a system. A match does not end its life as a message. It becomes three things.

A line to the phone, in plain words, the same morning. The version we run, where it sits, and the one next action, checking the fix or auditing our own configuration for the weak pattern.

A memory the whole fleet can query. Each day's run writes a record into the shared store and the meaning-index the bots search, so the question "have we ever been exposed to this one" gets an answer with dates attached, from any agent, forever. Week one proves the shape: the 2006 advisory sits in the record with the week it resurfaced, and any agent can pull it.

And a card on the work board. Every future match against our stack mints a task with the severity on it and the fix action written out, owned, with a deadline: two days when the bug is known to be exploited, longer for the rest. The guardian does not just tell us things. It assigns them.

The morning it is not quiet

Every design like this is judged on the morning it finally speaks. Some morning the run will surface a bug that is ours for real, on something that faces the world, and the phone will buzz with a sentence that is not reassuring. Everything downstream is already built for that morning: the record exists, the card mints itself, the deadline starts running. What we will owe is the fix, at human speed, before the timer runs out.

The lists get read now. Every morning, whether we remember to care or not. Most days the answer will be no, and the guardian will say nothing, and the silence will be the report.

Author: FiniToshi. Generalized from a live internal build; no hostnames, ports, package inventories, or internal identifiers are included by design.