Does Frontier AI Strengthen Cyber Defense? Evidence from Project Glasswing

arXiv:2611.06629 · econ.GN, cs.AI, q-fin.EC · Submitted 2026-11-04 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: I'm Ruth.

Warren: And I'm Warren, here with Iris.

Ruth: We're looking at the paper: "Does Frontier AI Strengthen Cyber Defense? Evidence from Project Glasswing".

Warren: Here's the gist of the paper. Project Glasswing, an initiative involving a small group of large US tech firms granted early access to Claude Mythos Preview, a model designed for finding software vulnerabilities, was used to study how access to frontier AI impacts defenders, external researchers, and attackers.

Ruth: The research yielded several key findings regarding the effects of this access.

Warren: First, the study provides evidence on how access to frontier AI affects firms' production of cybersecurity information.

Ruth: Access led to more vulnerability disclosures without any detectable deterioration in measured quality or a slowdown in remediation efforts.

Title and authors: Ruth: So we’re talking about this paper called "Does Frontier AI Strengthen Cyber Defense? Evidence from Project Glasswing." It looks at what happens when a few big US tech companies get early access to a model designed to find software flaws.

Warren: Right, it’s about how that access changes the game for defenders and attackers in the cybersecurity world. I’m curious how much of this effect is real when you look at the numbers they're throwing around.

Iris: Well, essentially, they set up a study where defenders and attackers are racing to find the same software flaws using different levels of AI access. The core idea is to see if having that early access actually translates into better security outcomes for those companies.

Ruth: Exactly, so they’re looking at how this access affects the production of vulnerability disclosures and how fast those flaws get fixed. They found that firms with this access doubled their disclosures compared to the Chinese and Russian tech firms without it, which is a big difference.

Warren: Doubled in what way? Did they find more flaws overall, or were they just disclosing them faster than before? Because that’s where I usually look when I’m thinking about strategy.

Iris: They found that treated firms disclosed about one fifty percent more high-impact vulnerabilities, and the interesting part is that they fixed reported flaws as quickly as before. The authors noted there was no slowdown in remediation efforts, which suggests the speed of fixing wasn't affected by just having access to this AI tool.

Ruth: That’s a key finding because it means the defensive benefit isn't just about finding more problems; it’s about being able to handle them with the same speed you already had. But they also looked at external researchers, and that part was quite telling.

Warren: External researchers who sell their findings through bug bounties, they reported half as many vulnerabilities in the code of the treated firms compared to other firms' code, without leaving the market entirely. That sounds like a strange trade-off for those researchers.

Iris: It suggests that when you have this better tool, internal teams are finding more flaws themselves, which means external researchers aren't competing as much with the internal teams on the same pool of vulnerabilities anymore. But they still found fewer findings in the treated firms' code overall.

Ruth: That leads into what they call an asymmetry in software security between different countries and firms, especially because they deliberately excluded companies based in China and Russia from this access study. That really highlights the geopolitical angle of Project Glasswing.

Title and authors: Warren: So, if you put that exclusion criteria into perspective, it functions a bit like an export control for this kind of frontier AI technology. It’s not just about the model itself, but who gets to use that advantage first and how much they can leverage it over others in the global software supply chain.

Iris: And when we look at the allocation of research effort, the paper suggests that when defenders and attackers search the same pool of flaws, a better search technology’s value depends less on just having a huge number of flaws found and more on finding them first.

Ruth: So this AI access let defenders remove latent flaws before attackers could even use them, which is significant because it shifts the advantage in time to their favor. The share exploited within a month of disclosure fell by about two thirds because a smaller share was already exploited upon disclosure.

Warren: That’s interesting—the total number of exploits didn't change, but the timing changed drastically. It means these additional disclosures helped defenders catch things before attackers could act on them, which is a defensive win.

Iris: But they also pointed out that researchers who got displaced by this access carried their effort to other codebases, meaning the security benefits of the model spill over to other parts of the software ecosystem through researchers’ reallocated effort.

Ruth: So we’re seeing this shift in how security is found and exploited, and it seems like a better search technology helps defenders win that race in real-time. But we need to keep an eye on where this advantage is being concentrated.

Warren: And that brings us into the improvements the authors suggest for their work, which are really focused on making sure the study captures these nuances more effectively. They say they could tune their system to prioritize finding high-impact, remotely exploitable flaws and flaws present in many different releases, because those types of vulnerabilities saw the largest increases in disclosures following access.

Iris: That makes sense; focusing on what matters most for immediate risk management rather than just the sheer volume of findings seems like a practical way to measure success here. It also mentions that they could focus on how much effort is spread across different research groups, which is important for understanding that spillover effect we talked about earlier.

Title and authors: Ruth: And from an economist's view, it’s about incentives—the model showed that access creates a situation where the race is decided by the relative capabilities of both defenders and attackers once comparable models are available to both sides. It’s not just who has the best defense, but how much capability you have compared to your opponent.

Warren: That points toward a future where security isn't just about building bigger walls, but about having superior search capabilities that allow you to react faster than the other side can exploit what you find. If everyone has this access, it sounds like the advantage shifts based on who is better at playing the search game.

Iris: And one of their main limitations is that they weren't able to fully track the timing evidence regarding Chinese vulnerability disclosure around when Project Glasswing launched; they found that timing measures there were weak, suggesting no discernible change in those specific disclosure patterns right after the launch.

Ruth: So while they found a lot about the access effect, their ability to precisely pinpoint immediate changes in specific geopolitical disclosure patterns right at the moment of launch was limited by the data they could gather.

Warren: It sounds like a good way to frame it: this study shows how AI access shifts the balance in finding and fixing flaws, but it also shows that measuring immediate geopolitical effects requires a much more granular time-series look than they were able to provide.

Iris: So, before we wrap up this segment on Project Glasswing, I just want to stress that the paper’s main takeaway is about the dynamic between defenders and attackers when both sides are using powerful AI models. It’s a study in relative capability.

Ruth: Absolutely, and that leads us into their final conclusion where they summarize how this access functions as a form of strategic technology allocation along geopolitical lines, making software security dependent not just on building the model but on access to it.

Warren: So if you’re driving home what this means for the average person listening, it’s that the security of your digital life is increasingly tied to which side of this new AI race you fall on geographically.

Iris: And so, to conclude on "Does Frontier AI Strengthen Cyber Defense? Evidence from Project Glasswing," the authors show that access to frontier models creates significant differences in vulnerability disclosure volume and speed, but also that the security benefits spill over through the broader research community.

Ruth: That’s it for this one. We’ve seen how this study on Project Glasswing shows that having access to frontier AI fundamentally alters the landscape of software flaw discovery and response, highlighting both defensive gains and geopolitical implications.

The paper's summary: Ruth: It turns out this study found that having that access didn't just mean finding more problems; it fundamentally altered the whole game of vulnerability disclosure.

Warren: Right, they showed that firms with this AI access doubled their disclosures compared to the Chinese and Russian tech firms who didn't have it. That’s a big jump in what information is being shared about software flaws.

Iris: And what’s interesting is that even though they disclosed more, the speed at which those companies fixed the reported flaws stayed exactly the same as before. No slowdown in remediation efforts was detected.

Ruth: So it suggests that having access to this frontier AI doesn't just give you a bigger list of problems; it allows you to process them at your normal pace.

Warren: That’s counter-intuitive, because usually, more flaws mean more work and slower patching. But the data shows they fixed reported flaws as quickly as before.

Iris: That’s the key part—the authors found that there was no detectable deterioration in the measured quality of those disclosures either. The speed and quality were maintained.

Ruth: And they looked at external researchers, which is where things get a bit strange for those who sell bug bounties. They reported about half as many vulnerabilities in the code of treated firms compared to other firms' code, even though they weren't leaving the market entirely.

Warren: So it seems that when you have this AI advantage, internal teams are finding more flaws themselves, which means external researchers aren't competing as much with those internal teams on the same pool of vulnerabilities.

Iris: That points to a reallocation effect; researchers who got displaced by this access carried their effort over to other codebases, spreading some of that security benefit into the rest of the software ecosystem.

Ruth: And then there’s that geopolitical element they built into the study by specifically excluding firms from China and Russia.

Warren: That exclusion criteria functions almost like an export control for this kind of frontier AI technology. It shows how access to these models becomes tied to geopolitical lines, not just technical capability.

Iris: The researchers also looked at how the search effort is allocated across different software vendors when both defenders and attackers are using similar levels of AI search technology.

Ruth: They found that when defenders and attackers search the same pool of flaws, a better search technology's value depends less on the sheer volume of flaws found and more on finding them first.

Warren: So this AI access lets defenders remove latent flaws before attackers can even use them, which shifts the advantage in time to their favor. The share of exploited vulnerabilities within a month of disclosure fell because smaller shares were already compromised when they were found.

Iris: That means these extra disclosures helped defenders catch things before attackers could act on them, but it’s not a guarantee that the total number of exploits went down overall.

Ruth: It’s more about the timing and the nature of the flaws being exploited immediately after disclosure. This study really shows that in this race, it's about who is faster at finding things first.

Warren: And for someone just listening to this, it means your digital security isn't just about building stronger walls; it’s about having superior search capabilities that let you react faster than the other side can exploit what you find.

Iris: Before we wrap up this look at Project Glasswing, keep in mind that the study itself had some limitations, specifically around timing evidence concerning Chinese vulnerability disclosure immediately after the launch.

Ruth: Exactly; they couldn't precisely pin down immediate shifts in those specific geopolitical patterns right at that exact moment. But the core finding about relative capability remains strong.

Warren: So we’ve seen how this access changes the landscape of flaw discovery and response, highlighting both defensive gains and geopolitical implications through this study.

The paper's improvements: Tom: So, the authors aren't just stopping there; they suggest how to make this research even better for future studies.

Ruth: They point out that they could tune their search system to specifically prioritize finding high-impact, remotely exploitable flaws and flaws that show up in many different software releases.

Warren: That makes sense from a risk management standpoint; focusing on what matters most for immediate threats rather than just the sheer volume of findings seems like a practical way to measure success here.

Iris: Yeah, because those types of vulnerabilities saw the largest increases in disclosures following the AI access, so targeting them helps isolate that effect.

Ruth: They also suggest focusing on how much effort is spread across different research groups when talking about that spillover effect we mentioned earlier.

Warren: That’s important for understanding how this model’s benefits actually get distributed throughout the whole software ecosystem, not just to the firms with direct access.

Iris: It sounds like a methodological tweak aimed at getting a clearer picture of where that security benefit is actually landing in practice.

Ruth: If you think about it, this shifts the focus from just counting total findings to assessing the value of those specific AI-discovered flaws for real-world defense.

Warren: It moves the study away from just measuring raw numbers and toward measuring actionable security risk reduction based on what’s actually exploitable.

Iris: That feels like a solid direction because it addresses the practical utility of the model in a real defense scenario, not just an academic curiosity.

Ruth: And from an economic standpoint, prioritizing those high-impact flaws aligns with how you would allocate resources when you have limited time and budget for patching.

Warren: It’s about optimizing that limited effort so it hits the most critical vulnerabilities first to avoid the biggest potential losses.

Iris: So the authors are essentially suggesting a way to refine the experimental setup to better isolate whether this AI access provides a tangible, prioritized benefit in an actual defense pipeline.

Ruth: That brings us back around to those initial geopolitical concerns, because refining these metrics helps you see if that strategic allocation is actually working as intended across different regions.

Warren: If you can measure the effect more precisely on high-impact flaws and the distribution of research effort, then maybe you get a clearer picture of whether this access is just a technical advantage or something bigger.

Conclusion: Tom: So we’re wrapping up Project Glasswing, which looks at how access to frontier AI impacts defenders and attackers in software security.

Ruth: We’ve seen how this study showed that having that access didn't just mean finding more problems; it fundamentally altered the whole game of vulnerability disclosure.

Warren: Right, they showed that firms with this AI access doubled their disclosures compared to the Chinese and Russian tech firms who didn't have it. That’s a big jump in what information is being shared about software flaws.

Iris: And what’s interesting is that even though they disclosed more, the speed at which those companies fixed the reported flaws stayed exactly the same as before. No slowdown in remediation efforts was detected.

Ruth: So it suggests that having access to this frontier AI doesn't just give you a bigger list of problems; it allows you to process them at your normal pace.

Warren: That’s counter-intuitive, because usually, more flaws mean more work and slower patching. But the data shows they fixed reported flaws as quickly as before.

Iris: That’s the key part—the authors found that there was no detectable deterioration in the measured quality of those disclosures either. The speed and quality were maintained.

Ruth: And they looked at external researchers, which is where things get a bit strange for those who sell bug bounties. They reported about half as many vulnerabilities in the code of treated firms compared to other firms' code, even though they weren't leaving the market entirely.

Warren: So it seems that when you have this AI advantage, internal teams are finding more flaws themselves, which means external researchers aren't competing as much with those internal teams on the same pool of vulnerabilities.

Iris: That points to a reallocation effect; researchers who got displaced by this access carried their effort over to other codebases, spreading some of that security benefit into the rest of the software ecosystem.

Ruth: And then there’s that geopolitical element they built into the study by specifically excluding firms from China and Russia.

Warren: That exclusion criteria functions almost like an export control for this kind of frontier AI technology. It shows how access to these models becomes tied to geopolitical lines, not just technical capability.

Iris: The researchers also looked at how the search effort is allocated across different software vendors when both defenders and attackers are using similar levels of AI search technology.

Ruth: They found that when defenders and attackers search the same pool of flaws, a better search technology's value depends less on the sheer volume of flaws found and more on finding them first.

Warren: So this AI access lets defenders remove latent flaws before attackers can even use them, which shifts the advantage in time to their favor. The share of exploited vulnerabilities within a month of disclosure fell because smaller shares were already compromised when they were found.

Iris: That means these extra disclosures helped defenders catch things before attackers could act on them, but it’s not a guarantee that the total number of exploits went down overall.

Ruth: It’s more about the timing and the nature of the flaws being exploited immediately after disclosure. This study really shows that in this race, it's about who is faster at finding things first.

Warren: And for someone just listening to this, it means your digital security isn't just about building stronger walls; it’s about having superior search capabilities that let you react faster than the other side can exploit what you find.

Iris: Before we wrap up this look at Project Glasswing, keep in mind that the study itself had some limitations, specifically around timing evidence concerning Chinese vulnerability disclosure immediately after the launch.

Ruth: Exactly; they couldn't precisely pin down immediate shifts in those specific geopolitical patterns right at that exact moment. But the core finding about relative capability remains strong.

Warren: So we’ve seen how this access changes the landscape of flaw discovery and response, highlighting both defensive gains and geopolitical implications through this study.

Iris: My final thought is that as long as these frontier models keep getting released, we need to keep watching how that strategic allocation of access shapes the global security landscape.

Thomas Mosk

Queen Mary University of London · Anthropic

econ.GN, cs.AI, q-fin.EC

Submitted: 2026-11-04

Updated: 2026-11-04

Project page: https://thomasmosk.com

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 100/100

The gist: Project Glasswing, an initiative involving a small group of large US tech firms granted early access to Claude Mythos Preview, a model designed for finding software vulnerabilities…

Key concepts

Frontier AI
This refers to advanced artificial intelligence models, like Claude Mythos Preview, that are at the cutting edge of capability. In this study, it was used specifically for its ability to search for software vulnerabilities in code.
Vulnerability Disclosure
This is when security researchers or firms publicly report a flaw they found in software. The study measured how much more flaws companies with AI access reported compared to those without, showing an increase in the volume of disclosed issues.
Asymmetry
This means there is an imbalance or difference in power between two parties. In this context, it describes the gap created when some firms have access to powerful AI tools for defense while others do not, creating a security advantage.
Precedence vs. Volume
This concept explains how the value of a superior search technology is determined. The research found that once both defenders and attackers have comparable models, the race to find flaws is decided by who finds them first, not just how many flaws are found overall.

Terminology

Summary

Project Glasswing, an initiative involving a small group of large US tech firms granted early access to Claude Mythos Preview, a model designed for finding software vulnerabilities, was used to study how access to frontier AI impacts defenders, external researchers, and attackers. p. 0 The research yielded several key findings regarding the effects of this access. p. 0

First, the study provides evidence on how access to frontier AI affects firms' production of cybersecurity information. Access led to more vulnerability disclosures without any detectable deterioration in measured quality or a slowdown in remediation efforts. p. 1 Specifically, firms with access more than doubled their disclosures relative to Chinese and Russian tech firms without access, and they disclosed about 150% more high-impact vulnerabilities. p. 0 Furthermore, treated firms fixed reported flaws as quickly as before. p. 0 External researchers who sell findings through bug bounties reported about half as many vulnerabilities in the code of treated firms compared to other firms' code, without leaving the market. Crucially, there was no evidence that attackers exploited these additional vulnerabilities; instead, the share exploited within a month of disclosure fell by about two-thirds because a smaller share was already exploited upon disclosure, while the total number of exploits did not change. p. 0 This suggests that frontier AI helped defenders find and fix more flaws before attackers could exploit them, but the restricted release creates an asymmetry in software security between firms in different countries. p. 0

Second, the research examines how AI-enabled vulnerability discovery affects the allocation of independent research effort across software vendors. p. 1 The results indicate that when defenders and attackers search the same stock of flaws, a better search technology's value depends less on the sheer number of flaws found and more on finding them first. Access to the model allowed defenders to remove latent flaws before attackers used them, and researchers displaced by this access carried their effort to other code, spreading some benefit beyond the firms with direct access. p. 0

Third, the study investigates the dual-use consequences of greater vulnerability disclosure. p. 1 The findings point to a general feature of AI in adversarial settings: when defenders and attackers search the same flaw pool, the value of superior search technology is determined by precedence rather than volume. p. 1 The evidence suggests that once comparable models are available to both sides, the race is decided by the relative capabilities of both defenders and attackers, not solely by defender capability. p. 5

The allocation of access also has a significant geopolitical dimension. p. 1 The exclusion criteria for Project Glasswing deliberately excluded firms based in or majority-owned by entities in China and Russia. p. 2 Relative to these excluded firms, the firms with access roughly doubled the latent flaws they removed from their software. p. 5 This restriction functions similarly to an export control, allocating strategic technology along geopolitical lines; if frontier models continue to be released first to selected nations, a country's software security will depend not only on its ability to build such models but also on its access to them.

In terms of timing evidence regarding Chinese vulnerability disclosure around the launch, timing measures—such as the median delay between CVE identifier reservation and publication in China versus other years—were deemed weak, suggesting no discernible change in disclosure patterns immediately following the launch. p. 2 However, firm-level analysis showed high shares of vulnerabilities with CVE identifiers (0. p. 978, 0.94, 0.95) and significant activity from independent researchers crediting vulnerabilities for Chinese software across various channels. p. 8 Furthermore, the disclosure gap analysis indicated that Glasswing × Post had a high proportion of vulnerabilities with a gap of at most 7 days (0. p. 2524), and the analysis showed that three months exploit code was 0. p. 4503 (with significance at the 1 percent level). p. 1

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Faster vulnerability discovery by leveraging frontier AI models for defensive work can more than double vulnerability disclosures relative to firms without access, leading to a surge in high-impact flaws found before attackers can exploit them.

  2. External researchers searching treated firms' code now compete with internal vendor teams using the model, resulting in external researchers reporting about half as many vulnerabilities in treated firms' code but more in other firms' code, without leaving the market.

  3. AI access allows defenders to find latent flaws that attackers were not using, which leads to a fall in the share of exploited vulnerabilities within a month of disclosure because the additional disclosures did not increase attacks: the share of the treated firms’ vulnerabilities exploited within one month fell by 0.9 percentage points relative to Chinese firms.

  4. The system can be tuned to prioritize finding high-impact, remotely exploitable flaws and flaws present in many releases, as these types of vulnerabilities saw the largest increases in disclosures following access.

  5. External researchers are reallocated to codebases where vendors lack frontier models, suggesting that the security benefits of the model spill over to the rest of the software ecosystem through researchers’ reallocated effort.

Sources

Related papers