In early August, cryptography professor Matthew Green sparked a heated debate across the cybersecurity community with a provocative thread on X and a detailed blog post. Green, a long-time observer of the tension between government surveillance and personal privacy, posed a startling question: What if AI makes software bugs so rare that law enforcement can no longer hack into criminals’ devices?
His argument challenges the current “uneasy truce” that has defined digital surveillance for the past decade. Instead of forcing tech companies to build backdoors into their products, governments have invested heavily in purchasing hacking tools and spyware that exploit security flaws, known as zero-days, to bypass encryption and access target devices. Green warns that the rise of AI could upend this entire arrangement, making AI government hacking increasingly difficult and potentially obsolete.
The Encryption Truce That Changed Surveillance
The debate around encryption and law enforcement access isn’t new. In 2014, then-FBI director James Comey popularized the concept of “going dark,” warning that widespread encryption would prevent authorities from accessing crucial evidence. Around that time, messaging apps like Signal, WhatsApp, and Apple’s iMessage rolled out end-to-end encryption, making traditional wiretapping nearly impossible. Apple also began encrypting data on its devices by default, protecting iPhones from unauthorized access.
Despite these warnings, authorities have continued to catch criminals through other means, including hacking into devices. This has worked largely because governments found a middle ground: rather than demanding backdoors that would weaken everyone’s security, they purchased sophisticated hacking tools from companies that discover and exploit vulnerabilities.
How AI Is Changing the Vulnerability Landscape
The core of Green’s argument rests on a simple premise: AI is becoming remarkably effective at finding security vulnerabilities. As large language models and machine learning systems improve, they can scan code faster and more comprehensively than human researchers, potentially identifying and patching bugs before they can be exploited.
This could make software significantly more secure, reducing the number of available vulnerabilities for AI government hacking tools. “I’m concerned that AI is going to make software much too secure,” Green wrote, predicting that companies will patch bugs at unprecedented rates as AI tools help them identify flaws more quickly.
Luna Tong, a researcher who has worked for prominent offensive security companies, agrees with this assessment. She describes the current situation as a “gold rush of bugs” but believes it’s a temporary phenomenon. As AI improves, she argues, these vulnerabilities will become scarce again.
Paolo Stagno, chief technology officer at Crowdfense, a company that develops and sells zero-day exploits to governments, acknowledges the potential disruption. He notes that while exploiting security flaws is currently the “most democratic system we have” for balancing privacy and surveillance, the status quo may not survive if AI eliminates too many bugs.
The Skeptics’ Perspective
Not everyone is convinced that AI government hacking will become impossible. Several researchers currently working in the offensive security industry disagree with Green’s assessment, offering compelling counterarguments.
Hamid Kashfi, founder of DarkCell and an AI cybersecurity startup executive, points out that AI’s impact cuts both ways. While it helps defenders find and patch bugs, it also assists offensive researchers in discovering more complex vulnerabilities. “For every AI found and reported bug out there, there are probably 20 that are not reported,” Kashfi explains, suggesting that many valuable vulnerabilities will remain hidden from vendors.
Other researchers argue that AI will primarily make easy bugs easier to find, while the more complex, valuable vulnerabilities that governments prize will remain accessible to skilled human researchers. They also point to modern security protections, which pose a greater challenge to hacking efforts than AI does.
Eva Galperin, director of cybersecurity at the Electronic Frontier Foundation, offers a nuanced view. She notes that AI is simultaneously creating and eliminating vulnerabilities. The rise of “vibe-coding”—developing with AI tools—may introduce new bugs even as AI helps find existing ones. Additionally, finding bugs doesn’t guarantee they’ll be patched quickly, as the patching process itself can be complex and time-consuming.
The Looming Backdoor Threat
Perhaps most concerning is what happens if AI truly makes hacking tools obsolete. Green warns that governments may renew their push for backdoors—intentional vulnerabilities built into devices and software for surveillance purposes. This would make everyone’s devices less secure by design, fundamentally changing the privacy landscape.
Katie Moussouris, founder of Luta Security and a veteran of bug disclosure programs, sees this as a genuine long-term threat. “We have some distance to go before the latest phones and laptops are completely bug free,” she acknowledges, but adds, “There will be some point at which finding bugs will be much harder and that may trigger these pressures to build in backdoors.”
Moussouris estimates that we have until after the next presidential election before the intelligence community feels materially hampered enough to push for backdoors seriously. This timeline suggests that while immediate disruption may be limited, the long-term implications for AI government hacking are significant.
What This Means for Privacy and Security
The debate around AI government hacking capabilities highlights a broader tension in digital security. For years, privacy advocates have fought against government demands for backdoors, arguing that such measures would weaken security for everyone. The current system, while imperfect, has provided a compromise: governments can access devices through exploits, but these tools are expensive, difficult to deploy, and limited in scope.
If AI eliminates this middle ground, we may face a stark choice: accept backdoors that weaken security for all users, or accept that law enforcement may not be able to access evidence on encrypted devices. Neither option is ideal, and the debate is likely to intensify as AI capabilities continue to improve.
As AI continues to advance, the cybersecurity community must grapple with these challenges proactively. Companies developing AI security tools should consider the broader implications of their work, including how it might disrupt the delicate balance between privacy and surveillance. Governments should invest in lawful access mechanisms that don’t rely on weakening encryption for everyone.
For now, the debate remains theoretical. AI hasn’t yet eliminated software bugs, and governments continue to purchase and deploy hacking tools. But Green’s warning serves as an important reminder that technological progress can have unintended consequences, and that the security solutions we develop today may create new problems tomorrow.
The future of surveillance and privacy may ultimately depend on how we manage these emerging challenges, ensuring that we protect both national security and individual privacy in an increasingly AI-driven world.
