Skip to content
AKNOSTIC
← Back to blog
Platform Engineering2026-11-16 · 5 min read · By Matthijs Bosman

CVE Sentinel, the early-warning channel

A Slack channel that alerts only when a CVE hits your stack and KEV, urgent wording or EPSS ≥ 0.30 signal exploitation.

We recently put a bot live that posts security news to one Slack channel. Its job is easy to describe: tell the team early when something we run is being exploited, and stay quiet the rest of the time.

Security channels that ping people for every advisory end up muted and alarm fatigue is real, so we would like to avoid that. We are still testing the bot and happy with the results so far but in early stages. This post is how it decides when to send an alert, and what it deliberately leaves to other tools.

Three levels, and only one of them pages

Every 15 minutes the bot pulls a set of feeds: disclosure lists, CERT and vendor advisories, researcher blogs, news sites and the CISA KEV catalogue. Everything it keeps lands in one of three levels:

  1. ALERT: the item matches our stack and has an exploitation signal (more on those below). This is the only level that sends @here.
  2. WATCH: the item matches our stack, but nothing says it is being exploited. It is posted without a mention.
  3. INFO: a new KEV entry for something we don't run. Also posted without a mention, so we still see what is being exploited elsewhere.

When we say the bot pages, we mean it posts with @here, so the people active in the channel get a notification. WATCH and INFO items are posted silently.

Everything else is dropped. A news item that names none of our stack terms and no CVE we already track never reaches the channel.

So nobody gets the alert unless two things are true at the same time: the news is about our stack, and there is a reason to think it is being exploited.

What counts as exploitation-grade

There are three signals, and any one of them turns a stack match into an ALERT.

The first is a KEV listing. CISA calls its Known Exploited Vulnerabilities catalogue "the authoritative source of vulnerabilities that have been exploited in the wild", and recommends that organisations use it as "an input to their vulnerability management prioritization framework". The bot reads the KEV JSON feed directly.

The second is wording. There is a short, editable list of urgent terms, such as "actively exploited", "out-of-band", "unauthenticated", "privilege escalation" and "zero-day". This one is about speed, because disclosure lists and news sites can say "actively exploited" before any catalogue has been updated.

The third is EPSS. FIRST's Exploit Prediction Scoring System "estimates the probability that a published CVE will be exploited in the wild in the next 30 days". It is a score between 0 and 1, published daily and free to use through a CSV download and an API. The bot pages when a CVE that matches our stack scores 0.30 or higher. That 0.30 is our own cut-off, not a FIRST recommendation. It is a tuning decision, and it lives in the same config file as everything else.

These signals answer different questions: KEV is observed exploitation, EPSS is a predicted likelihood. Neither is perfect on its own, but together with the wording, they are a practical way of saying "this one is worth interrupting people for".

The stack is the filter

The most important setting is the watchlist: the list of components we actually run. The README is very direct about it: "Stack terms matter more than anything else here."

Terms are matched case-insensitively as whole terms, so specific terms work best. "kernel" or "python" on their own are noisy, and a bare language name doesn't count as a stack term. Next to the watchlist there is a small baseline of things nearly every Linux estate runs, such as OpenSSH, OpenSSL, glibc and sudo, so an advisory about those gets through even when nobody listed them.

Every CVE gets one Slack thread, and later news about the same CVE goes into that thread. When a CVE we already track is added to KEV, or a thread jumps to ALERT, the bot posts that back to the channel with @here, so a story that gets worse doesn't stay hidden in a thread. The first run also seeds silently, "so the channel doesn't start with 500 messages", which would be a bad first impression for a channel you want people to trust.

Small on purpose

The whole thing is one Python script, one config file and one Slack channel. It runs as a Kubernetes CronJob every 15 minutes, and Flux syncs the manifests from Git, so nobody applies anything by hand.

Ten scenario tests cover failed fetches, rate limits and retries, and they never touch the network or Slack.

The first page

The first real stack match reached the channel on 5 November 2026, as an @here ALERT. It was a news story about Linux root escalation exploits. The stack match was Linux, and the wording "privilege escalation" turned it into an ALERT.

You can argue about whether that one deserved an @here. The page came from the wording, not from KEV or EPSS, and terms like "privilege escalation" and "unauthenticated" are broad, so they can page on news that isn't exploitation-grade at all. That is why the README has one clear line about them: "urgent_terms decides what pages people. Review it after the first weeks." We plan to do exactly that, and to tighten the list where it pages on noise.

What it deliberately doesn't do

The bot tells you what is happening in the world. It doesn't tell you where the affected software runs in your estate, so pair it with SBOM or node scanning.

The channel can tell you that a library is being exploited. It can't tell you which of your images ship that library, in which version, or on which nodes. That is what an SBOM and a scanner are for.

The bot's own image is a good example. A Trivy scan reports 45 HIGH and no CRITICAL findings for it (a list no news feed would have given us), and we accepted them with written reasons.

There is a smaller gap too: a news item that uses a name we didn't put on the watchlist is dropped, so an out-of-date watchlist means the bot stays quiet when it shouldn't.

We see the channel and the scanner as two parts of the same process. The channel tells you that something is worth looking at, and the scanner tells you where it runs.

Copying the idea

If you run a platform team, a narrow paging rule like this is easy to copy. Start with the watchlist, because everything else depends on it, and expect to tune the EPSS cut-off and the urgent terms along the way.

We are still testing the bot. After the test phase we may open-source it, and of course we can implement it in your infrastructure as well.

Interested in working together?

Let's talk