AIs ability to find major software bugs is growing 490% year on year
In the fast-evolving world of technology, a concerning trend is emerging: an unprecedented surge in software vulnerabilities discovered with the help of artificial intelligence (AI). This isn't just a slight increase; it's a "deluge" that is overwhelming tech companies, open-source projects, and security researchers alike. The sheer volume and, increasingly, the severity of these newly found flaws are posing significant challenges, straining resources, and forcing a re-evaluation of how software security is managed.
To understand the scale of this problem, consider the data from the Zero Day Initiative (ZDI). ZDI is the world's largest independent bug bounty program, which means it offers rewards to ethical hackers who find and report vulnerabilities in software from any vendor. This independence makes its data a good indicator of the broader cybersecurity landscape. According to information provided to Mashable, ZDI has witnessed an astonishing 490 percent increase in vulnerability submissions this month compared to the same period last year. This dramatic rise highlights a critical shift, and the month isn't even over yet.
The implications of such a massive influx are profound. "Organizations that receive bug reports are struggling to keep up with the triage and response process," explained Dustin Childs, Head of Threat Awareness at the Zero Day Initiative. Triage involves sorting through reports, confirming their validity, assessing their severity, and figuring out how to fix them. This process is time-consuming and requires highly skilled experts. When the number of reports explodes, even well-resourced teams can become overwhelmed.
The strain is so immense that some prominent programs have buckled under the pressure. Childs noted, "A couple of programs, most notably the Internet Bug Bounty program, completely shuttered their doors rather than try to keep up." The Internet Bug Bounty Program announced its closure on March 27, citing the crisis in bug submissions. Administered by HackerOne, a leading bug bounty platform, the program stated that this crisis was fundamentally changing the "landscape" of how bugs are discovered. The core reason? AI-assisted research.
AI tools are dramatically expanding both the "coverage" (the breadth of software being analyzed) and the "speed" (how quickly vulnerabilities are found) of vulnerability discovery. This efficiency, while seemingly beneficial for identifying weaknesses, is creating an unmanageable workload for the human teams responsible for patching these flaws. As AI technology continues to advance, it's not just finding more bugs; it's also uncovering much more severe vulnerabilities that demand immediate attention and patching. Cybersecurity experts widely believe that this deluge is only just beginning, with companies like Anthropic and others pushing the boundaries of AI capabilities.
The Claude Effect: AI's Potent New Frontier in Vulnerability Discovery
A significant factor contributing to this escalating crisis is the emergence of advanced AI models designed with powerful cybersecurity capabilities. Anthropic, a leading AI company, recently made headlines with the arrival of Claude Mythos. This AI model was touted as having "demonstrated a striking leap in cyber capabilities," to the extent that Anthropic claimed it was too dangerous for public release. The core claim was that Claude Mythos could autonomously discover and exploit "zero-day vulnerabilities" across all major operating systems.
What exactly are zero-day vulnerabilities? These are critical software flaws that are unknown to the software vendor or the public. This means there's no patch available to fix them, leaving systems completely exposed to attackers. Hackers who discover or purchase zero-day exploits have a significant advantage, as they can use these flaws to gain unauthorized access to systems or data without detection. Because they are unpatched and unknown, zero-days are considered the most urgent and dangerous type of bug, highly sought after by malicious actors and nation-state attackers.
Anthropic's strategy with Claude Mythos was to release it to a select, closed group of organizations, stating its intention to provide tech leaders with an opportunity to "secure the world's most critical software." The company further claimed that the AI found so many bugs that it was impossible to report them all at once. This approach, however, has drawn criticism from some cybersecurity experts who have dismissed it as "security theater and a publicity stunt." Critics argue that withholding information about such powerful capabilities, even with good intentions, could be problematic. Anthropic has, however, pledged to disclose all the vulnerabilities discovered by Claude Mythos once they have been patched by their respective maintainers.
Buried within its April 7 blog post about Claude Mythos, Anthropic included a stark statistic that underscores the scale of the problem: "fewer than 1% of the potential vulnerabilities we’ve discovered so far have been fully patched by their maintainers." This figure is incredibly telling. It implies that Claude Mythos has uncovered an immense number of potential security flaws, far more than the tech community can currently handle. The reason for this low patch rate, according to Anthropic, is that they meticulously triage (prioritize) these newly found zero-day bugs, disclosing only the highest-severity issues first. This measured approach is intended to prevent overwhelming organizations with an "unmanageable amount of new work."
Furthermore, Anthropic estimates that the discoveries made so far represent only "a small fraction" of the bugs it expects to find in the coming months. The sheer volume of vulnerabilities has already forced the company to hire dedicated security contractors specifically to assist with the complex and time-consuming disclosure process. This necessity highlights the significant resources required even for a well-funded AI company to manage the output of its own advanced models, signaling a future where human capacity will be increasingly stretched to keep pace with AI-driven discovery.
The Growing Volume and Severity of Bugs: A New Reality for Developers
Prior to the introduction of advanced models like Claude Mythos, cybersecurity researchers had already warned about a surge in bug reports fueled by earlier AI tools. However, a common complaint was that many of these AI-generated reports were of very low quality, often containing inaccurate or poorly researched information. While these low-quality reports didn't necessarily point to critical flaws, they still created a significant burden. As Dustin Childs explained, "Not every submission ends up being a real bug, but we still have to triage it as if it is." This means valuable time and resources are spent investigating reports that ultimately prove to be false positives or minor issues, diverting attention from genuine threats.
However, the landscape is shifting dramatically. The current trend indicates that the severity of bug reports is, once again, on the rise. This means AI is not just finding more bugs, but also more dangerous ones. This escalation in severity further complicates matters for developers and security teams, as critical flaws demand immediate, high-priority attention.
A prime example of this challenge comes from Daniel Stenberg, a renowned Swedish open-source coding expert and the lead developer of cURL. cURL is an incredibly widely used open-source project, a command-line tool and library for transferring data with URLs, embedded in countless applications and devices around the world. In January, Stenberg made the difficult decision to pause the cURL bug bounty program, directly attributing this action to the impact of AI. He revealed that cURL had received more bug reports in 2025 than in the previous two years combined, with projections indicating that this number would double again in 2026. This exponential growth rate is unsustainable for any project, especially one heavily reliant on volunteer contributions.
Stenberg openly discussed his rationale on his blog, stating, "The main goal with shutting down the bounty is to remove the incentive for people to submit crap and non-well-researched reports to us. AI-generated or not. The current torrent of submissions put a high load on the curl security team and this is an attempt to reduce the noise." His concern was the overwhelming "noise" – the sheer volume of reports, many of which lacked proper investigation and often didn't represent actual vulnerabilities. He aimed to alleviate the burden on his small, dedicated security team so they could focus on legitimate issues rather than sifting through AI-generated clutter.
Yet, in a critical reversal of last year's trend, Stenberg confirmed to Mashable that the latest surge of security reports does, in fact, represent genuine and significant security concerns. This is a crucial distinction: AI is now finding high-quality, actionable vulnerabilities. In an update this month, Stenberg wrote that he had heard from more than 20 open-source projects, "who all confirm this trend: a larger volume of decently high-quality security reports." This widespread confirmation from across the open-source community underscores that this is not an isolated incident but a systemic shift impacting the entire software ecosystem.
In his latest blog update, Stenberg further elaborated on the severity of the situation, confirming that both the volume of new bug reports and the severity of those bugs are increasing in 2026. He noted, "The rate of confirmed vulnerabilities is back to and even surpassing the 2024 pre-AI level, meaning somewhere in the 15-16% range." This figure indicates that a significant percentage of reports are now leading to confirmed, exploitable flaws, emphasizing the critical nature of these findings.
The human cost of this deluge is also a major concern for Stenberg. He worries about the impact on developers, particularly those working on volunteer-driven open-source projects. "I can only imagine that projects that are all volunteers, with a larger code base that perhaps has gotten less scrutiny, perhaps because they are younger, they can easily get drowned in quality reports," he stated. This "overloading" can take a severe "mental toll on many maintainers," potentially leading to burnout, project abandonment, and a decline in the overall health and security of essential open-source infrastructure upon which much of the digital world relies. If maintainers can't keep up, critical software components could remain vulnerable for extended periods, creating significant risks for end-users and businesses.
The question naturally arises: is this widespread zero-day deluge a direct "Claude Mythos effect" in action? While the timing aligns, and the capabilities of such models are evident, direct attribution remains challenging. Until Anthropic completes its comprehensive reporting on all the bugs Claude Mythos has discovered, it's hard to say for certain. Neither Childs nor Stenberg could definitively attribute the increases they observe to Mythos specifically. The reality is that multiple advanced AI models are likely contributing to this broader trend.
Indeed, there are also clear signs that private companies are experiencing a similar surge in AI-discovered bugs. Microsoft, for instance, announced a staggering 165 new bugs patched in its April security update. Childs highlighted this as "the second largest monthly release in Microsoft's history" in his Patch Tuesday blog, citing AI as a probable cause for the significant increase. Patch Tuesday is Microsoft's regular, monthly release of security fixes, and such a large number of patches is highly unusual. While Microsoft, in a statement to The Register, denied that AI was solely to blame for the unusually large update, they did credit Anthropic researchers for discovering one of the bugs. This acknowledgement, even for a single bug, underscores the reality that AI-driven discovery is now a recognized source of critical vulnerabilities across the industry.
Regardless of specific attribution to one AI model or another, the overall industry trend line is unmistakably clear: we are facing a massive increase in both the number of potential and confirmed security bugs, all of which require urgent fixing. This new reality demands a significant shift in how we approach software development and security.
AI and Cybersecurity: Navigating the Future of Digital Defense
The relationship between AI and cybersecurity is complex and multi-faceted. On one hand, AI tools offer immense potential to enhance our defenses. As stated in the Claude Mythos system card, Anthropic believes that AI tools will ultimately provide "more benefits to cybersecurity defenders in the long run." AI can automate threat detection, analyze vast amounts of data for anomalies, and even predict potential attack vectors, potentially strengthening our digital fortresses.
However, the short-term outlook presents a significant challenge. Cybersecurity experts, including those at Anthropic, acknowledge that hackers may have the advantage in the short-term. The very tools that can find vulnerabilities can also be weaponized. Anthropic itself notes that existing AI tools "already provide ‘significant help’ to the relevant threat actors in the sense of increasing their general productivity." This means malicious hackers can use AI to accelerate their reconnaissance, automate exploit generation, identify new attack surfaces, and even craft more convincing phishing attempts. This creates an immediate arms race where offensive capabilities might currently outpace defensive ones.
Paradoxically, AI is quickly becoming both the problem and the potential solution for developers. Faced with the overwhelming flood of AI-discovered bugs, developers and security teams are now turning to AI themselves to help manage the chaos. "We’ve begun using AI to aid in the triage process," Dustin Childs confirmed. "It’s the only way we’ll be able to keep up with this level of submissions." AI can be trained to help filter out "AI slop" – the low-quality, often repetitive or nonsensical reports generated by less sophisticated AI models – and prioritize genuine, high-severity vulnerabilities. Childs explained, "many entries are AI slop, but we’ve purchased a few of these bug [reports] just to teach our models what AI slop look like so we can avoid them in the future." This adaptive strategy is crucial for distinguishing actionable intelligence from mere noise, allowing human experts to focus their efforts where they are most needed.
The stakes are incredibly high. If the cybersecurity industry, including software developers, security researchers, and even AI companies, fails to adapt swiftly to this new reality, consumers will inevitably suffer the consequences. Unpatched vulnerabilities are an open door for cyberattacks, leading to data breaches, identity theft, financial losses, and a general erosion of trust in digital systems. "We’ve got to figure out how to scale up our fixes as fast as researchers (and attackers) are scaling up their findings," Childs emphasized. Without such rapid adaptation, users will have "little chance to apply these [fixes] in a timely manner," leaving them vulnerable to exploitation. This calls for unprecedented collaboration, innovative approaches to automated patching, and a fundamental shift in how we build, secure, and maintain software in an increasingly AI-driven world.
from Mashable
-via DynaSage
