Wired is reporting that even as artificial intelligence laboratories float the idea of an industry-wide agreement to slow the pace of AI development, widely available AI chatbots are already being used to surface security vulnerabilities at a scale that researchers and defenders are struggling to absorb.
To understand why this matters, it helps to step back from the dramatic framing of AI pause debates and look at what is quietly happening at the technical layer of the internet. Security research has always been a field defined by asymmetry. Attackers need to find only one exploitable flaw; defenders must account for all of them. For decades, the practical constraint on both sides was human labor. Finding vulnerabilities required skilled researchers with deep knowledge of specific codebases, operating systems, or network protocols. That constraint is eroding faster than most public discourse acknowledges.
Modern large language models, even the general-purpose chatbots available to ordinary consumers, have been trained on enormous volumes of code, documentation, vulnerability disclosures, and security research. That training means they can assist in identifying patterns that historically required considerable expertise to spot. A researcher who might have manually audited a few thousand lines of code in a week can now move through vastly more surface area with AI assistance. The problem is that this capability is not restricted to researchers working defensively. It is available to anyone with a subscription or, in many cases, for free.
This is not a theoretical concern sitting somewhere in the future. The conditions for a vulnerability explosion are already structural. The global codebase is enormous and largely unaudited. Open-source software underlies critical infrastructure, financial systems, and consumer devices at a scale that was not true even a decade ago. Legacy code written before modern security practices became standard persists in production environments everywhere. AI tools are now being pointed at all of it simultaneously, by people with intentions ranging from responsible disclosure to outright criminal exploitation.
The timing relative to the AI slowdown conversation is worth examining carefully. The pause or slowdown debate tends to center on frontier models, the most powerful and expensive systems developed by a small number of well-resourced laboratories. The implicit assumption in that framing is that the most dangerous AI capabilities reside at the frontier. The vulnerability situation complicates that picture considerably. The chatbots already in widespread use, the ones nobody is proposing to pause, appear sufficient to materially accelerate the discovery of security flaws. A governance conversation focused exclusively on the next generation of models risks missing the harm that current-generation models are enabling right now.
The consequences of this pattern are likely to fall unevenly. Large technology companies and well-funded enterprises have dedicated security teams and the resources to deploy AI-assisted defense at scale. They can, in principle, use the same tools offensively oriented actors are using, and do so faster. Smaller organizations, municipalities, hospitals, schools, and the vast middle layer of businesses that run on commercial software without large internal security functions, are far more exposed. They benefit from patches and disclosures but have little capacity to proactively hunt vulnerabilities in their own environments. The acceleration of vulnerability discovery widens the window during which unpatched flaws exist in the wild, and that window is most dangerous for organizations with the least capacity to respond quickly.
For software vendors, the pressure is also significant. If AI tools are genuinely accelerating the rate at which flaws are found, the cadence of patch development and deployment needs to accelerate with it. That is a resource and coordination challenge that does not have an obvious technical solution. Vulnerability disclosure norms, which have evolved slowly over years through negotiation between researchers and vendors, may need to be revisited in a world where the discovery rate has structurally increased.
There is also a subtler dynamic at work in how public attention is being allocated. The AI slowdown debate, whatever its merits in terms of long-term risk, commands considerable media and policy attention. It involves prominent figures, involves philosophical questions about existential risk, and maps onto existing political fault lines about regulation and technological progress. The vulnerability problem is less photogenic. It is incremental, distributed, and requires some technical literacy to appreciate. That mismatch in attention may mean that the nearer-term, more concrete risk continues to outpace the policy and institutional response directed at it.
What to watch for next is fairly clear. The volume and velocity of publicly disclosed vulnerabilities in the coming months will be a meaningful signal of whether the trend Wired is describing is accelerating. The behavior of major software vendors in terms of patch frequency and the handling of researcher disclosures will indicate whether the industry is adjusting to a faster tempo. And the degree to which AI safety policy discussions begin to incorporate current-model harms, rather than focusing almost entirely on hypothetical future systems, will determine whether governance catches up to the reality already unfolding.
