About this site
aisecurity.ninja is a curated reference for AI security: papers, tools, frameworks, courses, and communities, in one place. It began as my own working repository in early 2024 (when very few knew what AI security was, let alone practiced it) and became public because it seemed useful to other researchers.
It's a personal project, curated through an agent and kept free of hype. For more about my work, see sambucci.com.
Who maintains it
The map is kept by me, Luca Sambucci. I have worked in cybersecurity since 1991 and in AI since 2017. I quickly matched these two passions into AI security and never looked back. I am the founder of Noctive Security.
How the papers are sourced
The AI security papers page updates itself every day. I made an automated monitor that watches arXiv for new work on jailbreaks, prompt injection, abliteration, agent security, and related topics, writes a short summary for each, and delivers the new papers to this site through a secure ingest endpoint. Nothing is hand-typed; the archive grows on its own, newest first. Each paper links back to its arXiv abstract and PDF.
How the rest is kept current
The curated sections are reviewed on a schedule, every month. Each pass does the same four things: it checks that every link still resolves, it re-reads each entry's own page to see whether what we claim about it is still what it says, it looks for resources worth adding, and it stamps the review date you see at the top of each section. That date says when a section was last looked at, not when it last changed, because a section checked and found correct is exactly as current as one that needed an edit.
Entries are described plainly: what a resource is, who maintains it, and what it suits. Where a project has stopped moving, the entry says so with a date rather than a verdict, so you can judge whether that matters for your purpose. Inclusion is not an endorsement, and I avoid opinions about the worth of named products.
How I decide what is true
Every change needs a source and a quotation from it. A recollection that some tool is probably abandoned is exactly the kind of claim that rule exists to stop.
A dead link is not proof that a resource is gone, only that the address is. Before anything is removed I search for it under its own name, because projects move between organisations and guides get republished, and I request a page on the same site that I know exists, because some platforms return "not found" to any automated request while working perfectly in a browser. A replacement has to match the original on topic, scope and kind: a vendor's product page is never a successor to an educational guide. Where two sources disagree, the disagreement is described rather than quietly resolved in favour of whichever was read last.
Corrections
If something here is wrong, out of date, or should be added, write to legal-webREMOVE@toor.it, or reach me through sambucci.com. Corrections are acted on with evidence. An entry is amended or removed when a source shows it should be, and never on a single failed link. Every removal is recorded with its date, the address that died and the evidence for it, so a deletion can be reversed if it turns out to have been wrong.
How this site is funded
It isn't. This site carries no sponsored content.