I created CVE.radio to track published vulnerabilities, available fixes and known exploitation. Collection runs every thirty minutes and produces pages in French and English. A page is updated when sources report a new event: an available patch, reported exploitation or a revised advisory. Tracking continues after the CVE's initial publication.
It collects CERT-FR advisories and alerts, the CISA KEV catalogue, NVD data, GitHub advisories and Fortinet publications. Each CVE is then enriched using MITRE’s CVE Record, OSV, distribution trackers, FIRST’s EPSS score, CIRCL sightings and the Metasploit and Exploit-DB indexes. Connectors handle each source’s formats and rate limits.
The data is often incomplete: missing affected versions, scores not yet available, different product names across sources, or truncated paginated responses. The data model distinguishes a negative answer from unknown information. Every field retains its source. The presence of an exploit module is also distinguished from a confirmed ability to exploit the vulnerability.
AI analysis assesses the likelihood of automated exploitation of a CVE, based on the available information. This assessment complements patch information and known exploitation reports to help prioritise vulnerabilities.
Official French texts are kept as published. Other texts are translated by Mistral Small running locally. Checks verify identifiers, versions, URLs and CVSS vectors. A separate review model then compares the translation with the source text. Translations that fail these checks are withheld. A hash of the source text identifies descriptions changed after publication.
The generator produces static pages and JSON and XML feeds, published to a hardened front end. Visitor requests require no application processing. The same JSON format supplies the site, feeds and an MCP server for integrations with a CMDB, SIEM or agent.
Achievements
Updating CVE pages as new events are reported
Problem Vulnerability information changes after disclosure: a patch becomes available, exploitation is reported or an advisory is revised. The page needs to incorporate these changes to remain useful for tracking the vulnerability.
- - Linked new information to the relevant CVE, including CVEs already present in the catalogue.
- - Updated the page when sources report a new event or revise published data.
Result: Published CVEs continue to be updated from collected events. Readers can reassess priority and recommendations when new information becomes available.
Updating EPSS scores with each CVE collection run
Problem A CVE added after the EPSS file was downloaded could have no displayed score for up to 20 hours, even when the file already contained it. Analysts were therefore missing this input for vulnerability prioritisation. The cache delayed applying scores to new CVE pages.
- - Applied the latest available EPSS file on every collection run. Kept the download cache to limit network requests.
- - Updated scores for all CVEs in the database, including those added since the last download. Preserved modification dates to avoid unnecessary page regeneration.
Publishing CVE pages and JSON/XML feeds for monitoring tools
Problem Readers and monitoring tools need access to CVE pages and feeds during traffic spikes following major vulnerability disclosures. Access must not depend on the collector running or on queries to its database.
- - Generated HTML pages and JSON/XML feeds, then transferred them to the public server over rsync. Readers access the pages on the site; tools can retrieve the feeds without querying the collector.
- - Served by nginx with a CSP restricted to 'self', behind Varnish and Traefik.
- - Preserved official French texts. Other descriptions are translated with Mistral Small and checked by a separate review model before publication.
Handling incomplete vulnerability data
Problem Sources may omit affected versions or scores and use different product names. The JSON format returned booleans that conflated a negative search result with missing information. A client could conclude that no exploit existed without having the data to support that conclusion.
- - Added three states to the JSON format: true, false for an established negative result, and null for unknown information.
- - Defined negative results according to each source’s scope, treating absence from CISA KEV, absence of a public exploit and absence of exploitation reports separately.
- - Public tooling presence is measured against the complete Metasploit and Exploit-DB indexes, keeping existence apart from capability: an auxiliary module is a scanner and does not prove exploitation, and only an Exploit-DB entry marked verified allows saying it works.
- - Retained sources in each data block for verification and auditing.
- - Reported non-comparable versions, unknown vendor branches and ambiguous product names, with an explanation.
- - Used the same CVE format for public JSON, JSON-RPC and the MCP server.
Result: CMDB, SIEM and MCP integrations can distinguish negative results from unknown data and trace each field to its source.
- 1,116 'looked, found nothing' rows kept distinct from 89 'undetermined'
CVE translation with Mistral Small and verification by a second model
Problem A machine translation can alter an identifier or version number. It can also become outdated when the source description is revised after publication.
- - Kept the official French text where available, including CERT-FR advisories about a single CVE. Other descriptions are translated with Mistral Small.
- - A deterministic check runs before any model judgement: identifiers, versions, URLs, CVSS vectors and function names must form the same set on both sides, or the translation goes to review.
- - Compared the translation with the source text using a review model distinct from Mistral Small. The result determines whether the translation is published or queued for review.
- - Each translation keeps a hash of its source text. The NVD rewrites descriptions after publication: the translation then goes stale, stops being served and returns to the head of the queue.
Result: Mistral Small translations go through technical checks and the review model before publication. If a check fails, the page remains available without a French translation.