AI in cybersecurity: the original evidence study
- My contribution
- Coauthor: experiments, research, examples, formatting and edits. Sofia Plajutin wrote the initial draft.
- Scope
- April 2026 preprint · evidence audit, not a new capability benchmark.
- Evidence
- Original CSV artifacts and paper; public reproduction paths are incomplete.
Page by Julian Ramirez ·
A model finds a security bug. That is a meaningful result. It leaves several other questions unanswered: can someone reproduce it, does the finding survive triage, will a patch ship, and can the system stay within its authority while doing the work?
Those questions motivated the distinction at the center of Artificial Intelligence in Cybersecurity: evidence for a bounded task should not automatically become evidence for autonomous security operations. The original paper audits existing research and official sources rather than proposing a new benchmark.
Sofia Plajutin wrote the initial draft. I drove experiments, research and examples, and handled formatting and edits. This page explains the original study and its available artifacts. The paper is a preprint from April 2026; this project account was published on October 10, 2026.
Separate a source’s finding from our inference
One of the paper’s useful artifacts is a machine-readable evidence matrix. Each claim records its source and whether it is direct or inferred. That distinction makes the reasoning easier to challenge without pretending that every conclusion is a quotation from an authority.
Consider rows C04 and C06. The first records attribution in a Firefox advisory. The second interprets those credits as evidence that LLM-assisted discovery has entered vulnerability-disclosure workflows. They refer to the same source, but do different argumentative work.
Sources & implementation notes
Both rows point to mozilla2026mfsa13. C04 records advisory attribution. C06 is the authors’ synthesis, not a direct Mozilla statement.
That extra step matters. A reader can agree with the advisory’s contents while questioning how far our interpretation should travel. Keeping the inference visible makes that disagreement productive.
What the original evidence map supported
The original map groups the evidence into six task classes. Its source-access dates are in April 2026, and the assessment favors reviewed assistance, operator-plus-tools workflows and validated discovery over unbounded autonomous decisions.
This is a dated assessment, not a permanent ceiling on model capability. Its value is that the underlying claims and categories remain inspectable even when later evidence changes the picture.
The original CSV artifacts are available to read. The reproduction scripts reference manuscript, research and publishing paths absent from the public tree, so successful end-to-end reproduction is not promised by the repository as published.
What this research contributes
The contribution is an inspectable chain from source to claim to assessment. It lets a reader distinguish a reported result from our interpretation and question the inference without losing the original evidence. The six task classes organize that discussion; they do not establish a universal deployment policy.
Read the paper alongside its limits
The study supports a bounded reading of the available evidence. It does not establish that autonomous security operations are reliable, and the public artifacts do not make every reported result independently reproducible. Those limits belong beside the conclusions.
I discuss later evidence in AI security after discovery: capability is not deployment confidence. That October update is my separate interpretation, not a revision of the coauthored paper or an endorsement by Sofia.