Prioritization
Severity tells how bad a vulnerability could be. Most critical vulnerabilities are never exploited, and some medium ones are exploited widely. A project with a few hundred dependencies easily has dozens of findings, so severity alone does not say where to start. nanobom adds two public signals of actual exploitation to each finding that has a CVE, and lists findings by them first.
The signals
Known exploited (CISA KEV). The US Cybersecurity and Infrastructure Security Agency keeps the Known Exploited Vulnerabilities catalog: vulnerabilities with reliable evidence of exploitation in the wild. A finding with a CVE in it is known exploited, and shows:
- the date CISA added it;
- its due date, the deadline CISA sets for US federal agencies to remediate it;
- whether it is known to be used in ransomware campaigns.
Someone exploits these in the wild, which makes them the most urgent; CISA’s due date is a fair deadline for anyone.
EPSS. FIRST’s Exploit Prediction Scoring System estimates the probability that a CVE is exploited in the next 30 days, from 0 to 1, and its percentile among all scored CVEs. An EPSS of 0.0149 with a percentile of 0.78 means “1.5% chance, higher than 78% of CVEs”. It is a ranking aid, not a verdict.
Both are public data: CISA distributes the KEV catalog under the Creative Commons 0 1.0 license, and FIRST publishes EPSS freely, asking products that use it for credit (the interface says “EPSS scores by FIRST.org” on its About page, along with when each was last read).
Only CVE identifiers are looked up: those among the finding’s identifier and its aliases. A finding known only by a GHSA or RustSec identifier, with no CVE alias, has neither signal. When a finding has several CVEs, it takes the most alarming value of each signal: known exploited if any of them is, the earliest due date, the highest EPSS score.
Ordering
On the project page, findings are listed:
- known exploited first;
- then by severity, critical first;
- within a severity, by EPSS score, highest first; findings without a score last;
- then by CVSS score.
The upload response’s evaluation.findings is ordered by severity first (its 100 entries keep
the most severe), and within a severity in the order above.
The project page tags them KEV, counts them in its summary of open findings, and its filter
Only known exploited vulnerabilities keeps them alone; the organization page counts them per
project in its KEV column, and the nanobom_open_findings_known_exploited metric counts
them across the instance. A CI gate with fail_on=exploited fails on them whatever their
severity.
When the data is refreshed
After each successful run, nanobom refreshes the data of the CVEs of the run’s scope (those of open and ignored findings):
- the KEV catalog is downloaded and kept in memory for 12 hours, so most runs reuse it;
- EPSS scores are fetched, 100 CVEs per request, for the CVEs whose score is older than 24 hours.
The two feeds fail separately:
- when the KEV catalog cannot be downloaded, the copy in memory is used even if it is older than 12 hours, and without one (just after a restart) the KEV data stored for each CVE is kept as it was; EPSS scores are refreshed either way;
- when EPSS cannot be reached, the KEV data is still written from a freshly downloaded catalog,
so a newly exploited CVE is notified and fails
fail_on=exploitedgates during an EPSS outage; the EPSS scores stored for each CVE are kept as they were; - when both fail, nothing is written: the previous data is kept for every CVE.
The run itself is not affected by a failed refresh. Failures, including a failure of one feed
while the other was stored, are logged (refreshing exploitation data failed) and counted once
per refresh by nanobom_exploitation_refresh_failures_total.
Three bounds keep an outage of those services from slowing nanobom down:
- the KEV download and the EPSS requests each get 25 seconds, retries included; a feed that does not answer in time has failed, and the other is still refreshed and stored;
- a refresh taking more than 60 seconds is abandoned (and counted as failed), because uploads waiting for their evaluation are answered only after it;
- after a failed download of the KEV catalog, it is not asked again for 5 minutes, so a series of uploads during an outage does not wait for the retries each time.
The refresh comes before waiting uploads are answered so that fail_on=exploited sees the CVEs
first reported by that very run.
A vulnerability added to the KEV catalog after it was first reported is noticed at the next
refresh of a run covering it, at the latest the nightly run, and is notified once with the
exploited reason (see findings.new).
Limits
- The signals only cover CVEs; ecosystems whose advisories rarely get a CVE get few signals.
- EPSS predicts, it does not observe; KEV lists only what CISA has evidence for. Neither says whether the vulnerable code is reachable in your project: that is what triage records.
- Instances that cannot reach CISA and FIRST can use mirrors or turn the data off (Run without internet access); with it off, no finding is known exploited.