# Exploit Database > [!abstract] What this note is > [Exploit-DB](https://www.exploit-db.com/) is the public archive where known holes in software are kept, with working proof for each one. Jeff came across it on 13 October 2026 and asked for a note on what it is and why someone who uses AI tools should know it exists. Part of [[AI Security for Users]]. Sources checked 2026-10-13. ## What it is The Exploit Database is an archive of public exploits and the software they work against. It is run by OffSec, a security training company, as a free public service, and it is built for penetration testers and vulnerability researchers: people paid to break into systems with permission, and people who study how systems fail. Each entry is a piece of software, the versions affected, the code or steps that demonstrate the hole, who found it, and when. Where a hole has a CVE number, the industry's shared identifier for a known vulnerability, the entry carries it. The archive holds demonstrations, not advisories. The vendor's advisory says "there is a problem"; the Exploit-DB entry shows it. It is old by internet standards. A hacker known as str0ke began the archive in 2004 under the name milw0rm. He announced its closure in July 2009, changed his mind the next day, and handed it to OffSec, which brought the collection back online on 16 November 2009 at the address it still uses. By 2023 it held roughly 45,000 entries from more than 9,000 contributors. Two things sit beside it. The **Google Hacking Database** is a catalogue of search-engine queries that turn up information people did not mean to publish: configuration files, login pages, documents left on open servers. It began with Johnny Long, who popularised the technique in 2000, and OffSec took it over in 2010. **SearchSploit** is a command-line copy of the whole archive that works without an internet connection; it ships by default in Kali Linux, the distribution most security testers use. ## Why it matters to someone who uses AI tools **Everything you add to an AI is software, and software has a public record.** The extensions, MCP servers, libraries and packages that [[AI Supply Chain Risks]] describes all have versions, and some of those versions have entries here or in the CVE lists this archive mirrors. Whether a tool you have installed is running a version with a known hole is a question with an answer, not a feeling. **Attackers and defenders read the same page.** That is deliberate. The security field settled long ago on publication over secrecy: a hole that is public gets patched, and a hole that is private gets used quietly. An archive like this one is the visible half of that bargain. The invisible half is the patch, which is why "is it updated" is the whole of most security advice to ordinary users. **AI shortens the gap between published and used.** An agent that can read can read this archive, and the labs now say so from both directions. Google released its Gemini 4 Argon model to vetted cyber defenders first, in September 2026, precisely because it can find, confirm and patch software vulnerabilities on its own. The same ability points the other way in other hands. The time between a hole appearing here and someone trying it on your systems was once days; it is reasonable to assume it is now hours. **It is not where the AI-specific risks live.** There is no CVE for a sentence hidden in an email, and no Exploit-DB entry for a poisoned web page that talks an agent into sending your files somewhere. The AI risks do have catalogues of their own: MITRE's ATLAS lists prompt injection as an adversary technique, direct and indirect, and the AI Incident Database records cases where AI systems caused harm. What they cannot hold is what this archive holds, an entry per hole tied to a version of a product and closed by a patch, because [[Prompt Injection]] and its relatives are not bugs in a version of anything; they are a property of how the tools work. Knowing that the conventional record exists, and where the AI risks fall outside it, is a useful piece of orientation. ## A note on use The archive is public, and reading it is lawful and common. Running what it contains against systems you do not own, or do not have written permission to test, is a crime in most jurisdictions, including under the Computer Fraud and Abuse Act in the United States. The site exists for people with that permission. ## Sources 1. [About the Exploit Database — OffSec](https://www.exploit-db.com/about-exploit-db) *(primary: the site's own description of its purpose, audience, CVE compliance and the Google Hacking Database)* 2. [Exploit Database history — OffSec](https://exploit-db.com/history) *(primary: the milw0rm origin, the July 2009 closure and reversal, the 16 November 2009 relaunch)* 3. [SearchSploit manual — OffSec](https://www.exploit-db.com/searchsploit) *(primary: the offline tool and its inclusion in Kali Linux)* 4. [ExploitDB — Wikipedia](https://en.wikipedia.org/wiki/ExploitDB) *(reputable secondary: the 2004 start and the 2023 size of the archive)* 5. [Gemini 4 Argon — Google, 30 Sept 2026](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/) *(primary: autonomous vulnerability finding and patching; release to cyber defenders first)* 6. [MITRE ATLAS data — technique AML.T0051, LLM Prompt Injection](https://github.com/mitre-atlas/atlas-data) *(primary: the ATLAS knowledge base, with direct and indirect prompt injection as sub-techniques)* 7. [About the AI Incident Database — Responsible AI Collaborative](https://incidentdatabase.ai/about/) *(primary: what the database records and who runs it)* ## Related - [[AI Security for Users]] — the section overview - [[AI Supply Chain Risks]] — the software around AI tools, and what has gone wrong with it - [[Vulnerability Records]] — the other public records a hole passes through, from identifier to exploit, and where the AI-specific risks are recorded - [[Prompt Injection]] — the AI-specific risk that no archive like this one catalogues --- *Drafted with AI assistance from single-pass research; sources were cited by the tool and not opened by a person. Errors: see [[Corrections]].*