AI shifts zero-day economics to favor attackers
For years, zero-day vulnerabilities were treated as precious cyber weapons. They were rare, difficult to find, and required deep expertise in software behavior. A single zero-day could take weeks or months of manual work to discover, and nation-states guarded them as strategic assets. The trade in these flaws was secretive and expensive.
Today that scarcity has collapsed. Almost anyone can discover new zero-days on their own, using open-source AI models or proprietary systems like Claude and OpenAI. The same AI advances that help security teams write code and analyze logs can now search for flaws, chain vulnerabilities, and test exploit paths at machine speed. Attackers no longer need to manually inspect every dependency or reverse-engineer every application by hand.
The economics of zero-day exploitation have fundamentally changed. A human attacker could only pursue the most promising targets due to limited time and energy. An AI-assisted attacker can examine far more software, far more quickly, with far less fatigue. Even when most attempts fail, the sheer volume of attempts makes success more likely. In cybersecurity, scale transforms rare events into routine ones.
Modern software is too large and too interconnected for any organization to fully understand every weakness before an attacker does. Companies rely on a sprawling mix of commercial platforms, open-source libraries, cloud services, SaaS applications, APIs, identity systems, edge devices, and internal tools. Each layer introduces potential vulnerabilities. Each integration creates new attack paths. Each update silently changes the risk profile.
There is a massive amount of undiscovered vulnerability sitting inside software of all kinds. Some of it exists in old code. Some of it exists in new code written under pressure. Some of it comes from dependencies that few organizations track precisely. Some of it emerges from the way systems interact, even when each individual component appears secure. AI agents are well suited to explore that complexity because they can search broadly, persistently, and adaptively.
The Impossible Patching Problem
Defenders face an impossible patching problem. You cannot patch a vulnerability you do not know exists. You cannot prioritize every theoretical weakness equally. You cannot fully test every combination of software, configuration, and user behavior before deployment. Even mature vulnerability management programs are often built around known CVEs, vendor advisories, and threat intelligence feeds. Those are essential, but they arrive after discovery.
In a world of AI-accelerated zero-day hunting, discovery happens first in the hands of an attacker. This creates a growing game of whack-a-mole. A vulnerability is found. A patch is rushed. Exploitation shifts to a neighboring system, a different dependency, or a newly exposed interface. Security teams respond, but the backlog keeps expanding. AI increases the speed of the mallet on both sides.
Attackers benefit first because they only need to find one viable path. Defenders must protect all of them. The long-term answer may come from AI itself. As AI systems become better at generating, reviewing, and testing code, they should help eliminate entire classes of vulnerabilities before software reaches production. Secure-by-design development could become far more practical when AI can continuously inspect code, identify unsafe patterns, model exploitability, and recommend fixes in real time.
Eventually, common memory errors, injection flaws, authentication mistakes, and insecure configurations could be dramatically reduced because AI-assisted engineering catches them early. Organizations such as DeepTempo already report greater use of memory-safe languages such as Rust in their software. [1] While Rust is harder in some regards to code, it is safer, and coding it has become much easier thanks to AI.
A future where all software has been patched or rewritten safely is worth pursuing, but it is unclear how theoretically possible it really is for a variety of reasons. It has not arrived yet and will not arrive for many years. Today’s AI-generated code can still introduce vulnerabilities. AI security tools can still miss context. Enterprise environments still contain decades of accumulated technical debt.
Attackers are already using automation to move faster, while defenders are still integrating AI into existing workflows, approval chains, and risk models. Recent open-source benchmarks such as SOC Bench show that the best LLMs have a false positive rate in the 20% range. [2] At the scale of a typical large security environment, using these models would cost hundreds of millions of dollars a day. . The rise of AI currently vastly favors attackers.
Attack Resilience Over Patch Speed
Organizations must assume that some unknown vulnerabilities will be discovered by adversaries before they are disclosed publicly. Thanks to improved phishing and the widespread capturing of identities, attackers can increasingly simply log in. For the medium term, security strategy cannot depend entirely on prevention.
Prevention remains critical, but it must be paired with faster detection, stronger behavioral analytics, and tighter containment. Organizations including NIST, the Five Eyes, and the National Academy of Sciences have suggested that security teams should revisit approaches to understand what normal looks like across users, machines, identities, applications, and data flows
Zero-day exploits often succeed because they bypass known signatures, but they still leave footprints that today’s most powerful classification models can see and discern. Attacks may show unusual access patterns, abnormal privilege use, unexpected process activity, strange authentication flows, or data movement that, in combination, does not match the environment’s baseline. Purpose-built AI can help defenders find those signals sooner, especially when the attack itself is novel.
Companies should reduce the blast radius of compromise. Segmentation, least privilege, strong identity controls, continuous monitoring, and rapid isolation matter more when unknown vulnerabilities are being exploited. If an attacker uses a zero-day to gain entry, the next question is how far they can move before being detected. The smaller that window, the less valuable the exploit becomes.
Zero days have always rewarded speed, creativity, and asymmetry. AI agents amplify all three. The organizations that adapt will treat AI as more than a productivity layer for security teams. They will use purpose-built AI to detect unknown threats, prioritize weak signals that in combination with thousands of others suggest a compromise, and respond faster than human teams alone.
Sources
1. DeepTempo
2. NIST