PoC Monitor

Methodology

The PoC Monitor tracks, for every high-severity CVE, when it was published, when the first public proof-of-concept (PoC) appeared on GitHub, what that PoC sends over the wire, and when CISA added it to the Known Exploited Vulnerabilities (KEV) catalog. Every date shown comes from a stored source field. When we do not know a value, we show "unknown" rather than guess.

Which CVEs are tracked

CVEs published since 1 January 2025 with a CVSS base score of 7.0 or higher from any source (the CNA, CISA's ADP enrichment, or NVD), plus every CVE in CISA KEV. Rejected CVE records are not tracked.

t = 0: publication

t = 0 is the datePublished of the CVE record in the CVE Program's cvelistV5 repository. Events before t = 0 happened before public disclosure.

First public PoC

A public PoC is a GitHub repository that counts as a PoC (an exploit or a vulnerability checker that sends requests to a target) and is not suspicious, or an entry in Exploit-DB, a Metasploit module or a nuclei template. The first public PoC is the earliest date among them: a repository's GitHub creation date, an Exploit-DB entry's publication date, or the date a Metasploit module or nuclei template first landed in its official repository (the oldest commit touching the file, also under the old path when nuclei moved its templates). When none of them has a date yet, the CVE shows that a PoC exists but not when. A repository counts when our rules classify it as a PoC, or when PoC Index lists it and our rules cannot decide. A repository created before the CVE was published whose name does not carry the CVE ID (a collection of examples, a research harness) had the PoC added later, on a date we do not know, so it does not set the first-PoC date. We also record when we first found each repository. For CVEs published before the monitor existed, the creation date comes from GitHub, so it is not proof that we detected it then.

Repository creation is the best public signal we have. A repository can be created before its code is pushed, and history can be rewritten.

How we find PoCs

A crawler collects candidates every 30 minutes, newest CVEs first, without any language model:

Repositories that only index or aggregate PoCs for many CVEs (such as PoC Index itself) are never listed as a PoC. Each new repository is then judged by rules: it must be named for the CVE or present itself as a PoC for it, contain code or requests, and not be a list, write-up, detection or patch repository. Only cases the rules cannot decide go to a language model.

Suspicion scoring

Fake and malicious PoCs are common. We score every repository from 0 to 100 and list the reasons, for example: the owner account was created days before the repository, the README asks you to download a release, binary or password-protected archive, the repository contains executables or archives, the code decodes and executes obfuscated content or pipes a download into a shell, it references paste or exfiltration endpoints, or it never talks to a target at all. Repositories scoring 50 or more are marked suspicious. They are listed so you can avoid them, but they never count toward the first-PoC date, the characteristics or the detection rules.

We never run PoC code. Repositories are downloaded as archives into an isolated quarantine directory, only text files are kept, and they are analysed as text.

PoC characteristics

From each PoC we extract the requests it sends: method, path, query and body markers, notable headers, ports and User-Agent, each with a file and line reference. Parsers cover Python, Go and JavaScript HTTP clients, curl commands, raw HTTP requests in READMEs, and Nuclei templates. Lab and setup files (Docker files, setup, install and download scripts, CI) are never used: they describe the test environment, not the attack. Requests that are not the exploit, such as health checks, release downloads and calls to GitHub or IP lookup services, are dropped. A language model (Claude) only reads a repository, as untrusted data, when the parsers and rules cannot decide. The page shows the patterns shared across non-suspicious PoCs, ranked by how many repositories use them.

The Suricata and Sigma snippets are templates generated from these characteristics and should be reviewed before use. Regex rules marked "approved" were reviewed by a GriffinGuard analyst.

Edge devices

A CVE is tagged as an edge device when it affects an internet-facing appliance: a firewall or VPN, remote access gateway, router, load balancer, mail gateway, NAS or IP camera. The tag comes from a list of known products matched against the CVE record and its CPEs, and each tag shows why it was given.

Affected versions and mitigation

Affected versions, fixed releases, the vendor's fix and workaround text, and advisory links come from the CVE record as published. When the record names fixed releases but no fix text, we write "Upgrade to … or later" from those releases. When the record names no fixed release, we add the affected ranges and patched versions from reviewed GitHub Security Advisories that list the CVE, marked with the advisory ID; when the record does name one, the record is used alone.

CISA KEV

The KEV date is the catalog's dateAdded. KEV lists a CVE once CISA has evidence of exploitation, so it often follows the first PoC by days or weeks.

Data sources

Machine-readable data: the JSON API.