From https://blog.mozilla.org/en/privacy-security/ai-security-zero-day-vulnerabilities/:
Encouragingly, we also haven’t seen any bugs that couldn’t have been found by an elite human researcher. Some commentators predict that future AI models will unearth entirely new forms of vulnerabilities that defy our current comprehension, but we don’t think so. Software like Firefox is designed in a modular way for humans to be able to reason about its correctness. It is complex, but not arbitrarily complex.
By 2029, will this turn out to be false?
Update 2026-05-01 (PST) (AI summary of creator comment): The market requires that AI finds vulnerabilities that elite human researchers could not have found, not merely vulnerabilities that they have not found. A vulnerability going undiscovered does not qualify if it was theoretically within human capability to find.
Update 2026-05-01 (PST) (AI summary of creator comment): A vulnerability counts if it would be realistically impossible for a human to discover (e.g., because of the amount of context they would need to hold in their head to notice it).
Update 2026-05-01 (PST) (AI summary of creator comment): A vulnerability qualifies if a human could not see it even when looking at the right place, because the amount of context needed to hold in one's head (e.g., various interacting functions) is too great. This is distinct from vulnerabilities that humans simply have not found — the key criterion is whether it was realistically impossible for a human to discover, not merely undiscovered.
Update 2026-05-01 (PST) (AI summary of creator comment): The creator will subjectively decide the resolution, possibly deferring to the consensus of security people. The goal is to track whether the quoted claim from Mozilla will turn out to be false.
Update 2026-05-02 (PST) (AI summary of creator comment): Vulnerabilities in AI systems are out of scope (e.g., finding adversarial inputs with AI does not count). The market is specifically about human-written/human-maintained code.
Update 2026-05-09 (PST) (AI summary of creator comment): In some situations, the creator will resolve based on consensus of cybersecurity researchers. Clear resolution scenarios include:
Resolves No: consensus among cybersecurity researchers that AI did not find such vulnerabilities
Resolves Yes: consensus among cybersecurity researchers that a specific vulnerability could not have been found without AI
Some cases may require subjective judgment by the creator.
1,000People are also trading
If you don’t have any cybersecurity background, I ask that you don’t ask clarifying questions (and if you’re uncertain about something, don’t trade on this market).
What about vulnerabilities in deep learning systems? E.g. the adversarial case found that defeated KataGo and imperceptible changes to images that defeat classification neural networks. Both of those vulnerabilities were found themselves by deep learning methods and I don't think they could have reasonably be found by humans.
If you're specifically talking about vulnerabilities in human-written (or at least human-maintained code) I don't think this will happen, because humans need to understand the systems in order to maintain them. But a system beyond full human comprehension could indeed have vulnerabilities that humans could not have found.
@PlasmaPower try to think whether Mozilla would consider it to be a relevant kind of vulnerability.
Do you have a cybersecurity background?
@ms Yes I do have a cybersecurity background. I actually got most of my mana from finding an XSS in Manifold. I asked because I think Mozilla is only concerned with vulnerabilities in their products and supply chain and such, whereas I assume that you intend to also include other pieces of software, and I wasn't sure where that boundary was. E.g. you mentioned rowhammer in another comment, so maybe hardware vulnerabilities also count?
@PlasmaPower thanks, that clarifies it! It doesn’t have to be in Firefox; but I think vulnerabilities in AI systems are outside the scope. (Eg finding adversarial inputs with AI is not a relevant question.) I’m talking about human-written/human-maintained code
@ms People have already traded this market before this addendum, and there is a lack of real resolution criteria beyond vibes (based on my interpretation of the conversation below). Unless you would like to buy every non-cybersecurity expert trader and me out of our positions on the market, I think it is reasonable for traders to ask questions.
@prismatic why did you open a position before asking questions if the resolution criteria were unclear to you?
@ms It was clear at the time, then there was more discussion in the comments, and now it's not clear. Plus, we aren't allowed to ask questions anymore.
@prismatic what’s a comment that made it unclear compared to what you thought before?
Someone without a relevant background asked annoying questions ; I might not answer more of those, but feel free to try asking.
AI recently found a vulnerability: copy.fail.
This was technically AI-assisted, but the team acknowledged that the vulnerability itself was found by the AI. This affects all linux kernels from 2017. In 9 years, the most elite human researchers have not found this exploit used in one of the most used systems in the world.
@ms Then, could you please provide an example of what hypothetical situation you would resolve yes for?
@Air e.g., because of the amount of context that one would need to hold in the head to notice it’s a vulnerability
@ms please explain then? Context is a very subjective measure here especially because you haven’t really elaborated on your definition.
For some vulnerabilities I imagine exist, you cannot see them as a human even if you look at the right place because the amount of context you need to hold in your head (various functions etc.) to see one is too much.
This is unrelated to the amount of code unrelated to the vulnerability itself that one might be scanning.
@ms I understand what you saying but then what the actual resolution criteria? Someone publishes a substack explaining "how awsome this was"? A computerphile episode?
@Air does your argument prove that all of the previous Linux kernel vulnerabilities that were discovered by humans were impossible for humans to discover?
@hidetzugu I’ll subjectively decide, possibly deferring to the consensus of security people. The idea is to track whether the quote will turn out to be false.
@ms i’m confused as to what you mean by that. Does your reply mean that you’re looking for a type of exploit that has never been seen before? Something so severe that none of the hundreds of thousands of CVEs have never even touched?
@ms no, but I shouldn’t need one to ask why a recently discovered exploit is disqualified, nor should I need one to ask for clarification?
@Air sorry, I suggest you don’t trade on this market.
The question isn’t about severity of the vulnerability; it’s about whether they’re of the kind that people can, in principle, given enough time, comprehend/discover.
You’ve not provide any evidence of this being the case with this particular vulnerability.