AI bug-hunting could reopen the encryption backdoor fight

For more than a decade, U.S. law enforcement and intelligence agencies have lived with a quiet compromise: instead of demanding that Apple, Signal, and WhatsApp weaken encryption for everyone, the government buys the bugs that let them break into a specific phone. That workaround has a problem now, and a Johns Hopkins cryptographer says AI is what made it.

In a blog post and X thread on August 14, Matthew Green, who teaches computer science at Johns Hopkins and helped design the Zerocash protocol, argued that large language models are about to find software vulnerabilities faster than anyone can stockpile them. "I'm concerned that AI is going to make software much too secure," he wrote, and the worry is not that defenders win. It is that the loss of the bug supply will push governments to reopen the encryption backdoor debate they have been careful to keep on a shelf since 2014.

To understand the stakes, it helps to remember the fight that almost happened. In 2014, then-FBI director James Comey popularized the phrase "going dark" to describe a world in which encryption made criminal suspects unreachable. Around the same time, Apple's iMessage, then Signal and WhatsApp, rolled out end-to-end encryption to ordinary users; Apple later made the data on its phones encrypted by default, raising the cost of cracking any device protected by a strong passcode. Civil liberties groups and most of the cryptography community warned that any "golden key" for law enforcement would also be a key for spies, criminals, and foreign governments. Silicon Valley and Washington stopped short of a legislative collision.

What replaced the fight was what Green calls an "uneasy truce." The FBI and its foreign counterparts gave up on mandated backdoors and turned to a parallel market: buying exploits and spyware from private vendors such as NSO Group and the Italian firm Hacking Team, often through brokers who specialize in zero-day vulnerabilities, the kind of unknown software flaw that has no patch yet. The deal held because the supply of bugs was large enough, and the price was low enough, that agencies could keep getting in without weakening the encryption that protected everyone else. As Green put it, governments had stopped asking for keys and started picking locks, and the locks kept being breakable.

His provocation is that AI collapses the supply side of that arrangement. The same large language models that can write code and summarize legal documents are getting noticeably better at static analysis and fuzzing on the low-level C and C++ code where memory-corruption bugs live. If a frontier model can find a critical vulnerability in days rather than months, defenders can patch it in days too. Exploit brokers, who price bugs based on rarity and window of exposure, lose leverage. Intelligence agencies, who buy on that same market, lose access.

The second-order effect, and the part that worries Green most, is political. If the bug supply genuinely shrinks, the cost-benefit math that has kept the backdoor debate dormant starts to flip. A lawmaker who can no longer get a warrantful hack through a vendor can be tempted to ask why the underlying encryption exists in the first place. The result would be the fight encryption advocates have spent a decade dodging: a renewed push for client-side scanning, lawful-access mandates, or some version of the "responsible encryption" proposals that have circulated since the Obama and Trump administrations. TechCrunch's write-up of Green's argument surfaced this week and is what pulled it back into the press.

There is room to push back on Green's timeline. The going-dark argument has been contested since 2014: the UK's GCHQ published a so-called "ghost proposal" in 2018 that privacy researchers largely tore apart, the FBI's 2016 fight with Apple over the San Bernardino iPhone ended with the bureau buying its way in, and the courts have repeatedly declined to treat encryption as a thing companies can be ordered to break. AI-assisted bug discovery is also an empirical question, not a settled fact; researchers have published mixed results on whether AI models reliably beat human-led fuzzing on real-world code, and the productivity claims Green leans on are early rather than established. The spyware workaround has its own human-rights cost that this argument tends to underweight.

The reason the argument still matters is that the policy menu is not symmetric. If Green is right, governments either lose a capability they have come to depend on, or they get the keys they have spent a decade asking for. Either outcome lands on the same phones, the same Signal threads, and the same journalists' devices. The next round of the encryption fight is being set up now, and the people who will live with the result are not the ones buying the exploits.