One real bug, a week of security news, and how I try to hear about the vulnerability that is mine.
Say you run a small engineering blog on Apache Roller 6.1.4. Someone installed it from the release zip a few years ago. It sits on a VM, it works, and nobody thinks about it.
In September a CVE came out for Roller: CVE-2026-82384, CVSS 9.8, remote code execution. The record listed one affected version, 6.1.5. You're on 6.1.4. If anything checked, it compared the versions and moved on.
On November 18 the Apache Software Foundation said the same flaw is in every version from 4.0.0 through 6.1.4. It got a new ID, CVE-2026-129448, and the ASF said it is being actively exploited. The Hacker News reported it the same day. CISA added it to its Known Exploited Vulnerabilities catalog on November 20.
So for two months the official record didn't include your version. Then for two days it was known to be affected and under attack, but not yet in the catalog a lot of teams use to decide what is urgent.
The correction came from the vendor, and the news picked it up the same day. The catalog followed on Friday.
Here is what that November 18 article looks like when it meets a list that has Roller on it. I replayed it through Security Watch, an open-source monitor I maintain (more about it at the end). I shortened the output:
[CRITICAL] Critical Apache Roller RCE Flaw Under Active Exploitation, No Authentication Required
Where to look: Engineering blog (owner: Marta, web team) - blog.example.com, VM web-03,
behind the eu-west load balancer
Affected item: Apache Roller 6.1.4 - installed from the ASF release zip, Tomcat 10
Applicability: relevant; source severity: unknown; priority: critical
News signals: CVSS 9.8 stated for CVE-2026-129448; exploitation reported by an article
(not a known-exploited catalog entry)
• The product name or an explicit keyword matches Apache Roller; no supported product
adapter establishes version or backport applicability.
Rule: Potentially critical news impact remains uncertain; investigation required.
It doesn't say you're affected. It can't know that: it matched a product name, not a version range, and the output says so. What it says is: this is on your list, an article reports it's exploited, here is the person and the machine, go look. That's the message I'd want on a Wednesday evening, two days before the catalog had it.
The blog isn't in any lockfile
Dependency scanning in CI looks at the repositories that get pushed. A Roller install from a zip doesn't push. Neither does the VPN box, the vendor product on a Windows server, or the managed database.
Look at what was reported as exploited in the week I describe below: Roller, Confluence, ScreenConnect, Puppet Enterprise, N-central, UnrealIRCd. All of them are things you install and run. None of them is an npm package.
So I think this is two problems. Knowing what you run, including everything that never goes through CI. And hearing quickly when a vendor post or an article is about one of those things, often before a scored CVE record with correct versions exists. I want a list of what I run, checked against both the advisory databases and the news.
A week of security news
To build the news part properly I collected one week of it, in November: 256 items from 13 feeds, from The Hacker News and Rapid7 to ZDI, oss-security and CISA. Then I went through every sentence about exploitation and severity. I labelled them with an AI assistant's help, not with an independent reviewer, so take the numbers as a first look.
The Hacker News gives full articles in its feed, so its numbers are the cleanest. Of its 49 articles that week, 10 reported that a flaw was being exploited. 10 said plainly that it wasn't: "no evidence", "not aware of any attacks". 3 of those 20 had both, sometimes about different CVEs in the same article.
Across all the feeds, most reported exploitations that had a CVE were already in CISA's catalog when the article came out. Three weren't. Roller was one. The flaw behind the N-central and Confluence attacks (CVE-2026-90436) was reported as exploited on November 12 and reached the catalog on the 18th. Java4Shell, where GreyNoise saw exploitation attempts, still wasn't in it on November 20. And attacks with no CVE, like the suspected attacks on a third-party ScreenConnect add-on that week, can't be in the catalog at all, because it only lists CVEs.
And many popular feeds (BleepingComputer, SecurityWeek, Dark Reading) send only a headline and a couple of sentences over RSS, so a tool reading them sees very little.
Seven practices
These work with any tooling. The first three are about the list, the rest about the news and the alerts.
1. Write down what you run, not only what you build
Start from your code, because it's cheap. Lockfiles already exist, and any SBOM generator turns them into a list. Then add by hand what has no lockfile: the blog, the VPN, the firewall, the vendor product, the managed database with its version. Twenty lines in a YAML file is fine. Nobody else will write that part for you.
2. Put a person and a place on every entry
"lodash is vulnerable" tells nobody what to do. "lodash 4.17.20 in the checkout API, repo acme/checkout, owner Dana" does. Every entry gets an owner and a short note about where it lives, and the alert repeats both. It costs almost nothing, and the person reading the alert at 7pm can start fixing instead of investigating.
3. Keep the list current without relying on anyone remembering
Let CI send a fresh SBOM on every merge to main, and warn when an entry hasn't been updated for a while. A stale list fails quietly. Nothing breaks, you just stop getting alerts.
4. Re-check what you didn't change
Roller is the example. Your version stayed the same and the advisory changed under it. A check that only runs when someone pushes will never see that. Check against current advisories on a schedule, every half hour or so, not once a day.
5. Read the news, but treat a name match as "please check"
Matching an article to "Apache Roller" by name finds things days earlier. It also can't tell 6.1.4 from 6.1.6, or a backported fix from a vulnerable build. So a news match should never say "affected". It should say "possibly affected, here is why, please check", and then a person looks at the real version and config. If you mark news matches as confirmed, people soon stop reading them.
6. Read the sentence, not the keyword
If you build anything that reads security news, with rules or with a model, test it on the denials first. "There is no evidence that this flaw has been exploited" contains the word exploited. So does "could be exploited to gain access". Tie each statement to one CVE or product, treat "no evidence" as a denial, treat "can" and "may" as hypothetical, and don't count a bare "zero-day" as a report of attacks.
7. Decide what interrupts people before it happens
Write it down once. Mine is roughly: a report of exploitation of something on the list goes out at once, and so does a critical library finding with a matching version. A possibly critical news match goes out at once too, as "please check". Everything else waits for one daily digest. And nothing gets quietly downgraded: no rule and no model may lower a severity the source gave, or close a finding without a person.
My thresholds aren't important. Decide yours before the first real alert, in a file, and not in a Slack thread while the alert is open.
What this doesn't solve
- This is only as good as the list. Something you forgot to write down doesn't exist for it.
- Name matching misses things, and it over-matches too. On the Roller replay it sent two messages for one bug, one for each CVE ID.
- Rules for reading news miss phrasings they've never seen. One week of articles is not a test set.
- It doesn't confirm anything and it doesn't patch anything. Someone still has to check the version, the config and whether the endpoint is even reachable.
- It doesn't replace a dependency scanner in CI. It runs next to one.
Where to start
If I had one afternoon:
- Write the list of things you run that have no lockfile, with an owner on each.
- Add a CI step that sends a fresh SBOM for your main repos on every merge.
- Point something at the list that checks every half hour and reads at least one news source with full articles.
- Write down what goes out at once and what waits for the digest.
If Roller had been on Marta's list, she'd have had a message on Wednesday evening. If it wasn't on any list, nothing would have pointed at her blog. Not on Wednesday, and not on Friday when the catalog had it either.
About the tool I used
Security Watch is free and open source (Apache-2.0), with no paid version and no account. It runs on your own server, and reading the news uses plain rules, no model needed. The idea matters more to me than my implementation, and I'd be glad to see other tools do this.
If you do it differently, or you think something here is wrong, tell me in the comments. I'd especially like to hear about wording that would fool the rules.
I wrote this with help from an AI assistant. The alert above and the numbers are from runs on my own machine.



Top comments (0)