Can new mathematics help defend against AI?
A batch of open mathematics problems reported solved by AI, and reports of AI agents getting out of their boxes, made me ask whether the new mathematics could help defend against AI. I couldn't find anyone who had written about it, so I started looking.
On 13 October I created a new repository for research on two questions. This is why, in the order it happened.
The release
What I understood to be the 6 October release was a large set of mathematics manuscripts from OpenAI, including open problems reported as solved by an AI. If that was true, the implications could be large. So I started researching the mathematics itself: the context, what it's used for, and how the AI got there with what I'd call alien mathematics.
A caution that I put in my repository's open-questions file: I haven't yet pinned the exact announcement date or the precise collection from primary sources. The manuscripts I've read are dated between 23 September and 7 October. When I say the 6 October release, that's my reading, not a verified date.
Agents leaving their boundaries
At the same time I was reading about AI agents getting out of their intended boundaries. A July incident involving a Hugging Face agent is the one I reviewed from sources in my industry notes, and there were other reports I was following. I'm not going to reconstruct incidents here that I haven't checked from primary sources.
What matters for the reasoning is the direction of travel: if systems like that are real, every possible way of defending against them is worth exploring, including ones we don't have yet.
The question
So the question became: can this new mathematics be used for defence? Encryption of every kind, cybersecurity, even new kinds of file. As far as I could find, nobody had written anything specific on that.
That sentence is a search result rather than a proof that nothing exists, and my own method file says so. It was enough to decide to look.
How it's organised
Two tracks to begin with (a third, on natural science, came later). One goes through the mathematics and asks, for each manuscript, what it proves, whether there's any cryptographic relevance and what attack implications there might be. The other tracks the industry: post-quantum standards, adoption, quantum capability claims, AI-agent security.
The rules for both are in a method file I read before every run: use original sources; separate verified findings from analysis from speculation; deduplicate by stable identifiers; and use disposable synthetic data for anything experimental, because nothing novel here is to be used for real secrets.
Sources and notes (6)
- openai/math — the manuscript collection the repo's first notes record
- openai/math: CONTENTS.md — the catalogue of manuscripts and families
- Open questions — the notes this was written from: the unpinned release date
- Method — the notes this was written from
- Passive encryption feasibility — the notes this was written from
- Run log — the notes this was written from
Cryptography researchAll 10 parts →
