The End of Secrets

For most of human history, keeping a secret was a fairly straightforward business. You whispered it to someone you trusted, burned the letter, or buried the evidence beneath a loose floorboard. A sufficiently determined government could still discover what you were up to, but it had to employ spies, intercept couriers, or send a large man to kick down your door. Surveillance was expensive, intrusive, and constrained by the ordinary physical limitations of the world. The government couldn’t be everywhere, which was fortunate, since it generally wanted to be.

The internet changed all this, although perhaps not in quite the way its early enthusiasts expected. It gave the state an unprecedented capacity to observe its citizens, while also producing a peculiar mathematical defense against that observation. Encryption allowed ordinary people to communicate across the planet through channels that were entirely public, without permitting anyone else to understand what they were saying. An intelligence agency could intercept your messages, but provided the mathematics held and the system was properly implemented, it might as well have been collecting Egyptian hieroglyphics.

Now artificial intelligence is beginning to raise an uncomfortable possibility: What if the mathematics doesn’t hold?

In a recent essay for AI Frontiers, AI researcher Govind Pimpale examines the possibility that increasingly capable AI systems could discover fundamental weaknesses in public-key cryptography, the technology that makes secure communication between strangers possible. His argument rests on an important distinction. Modern public-key systems rely on mathematical problems that appear easy to solve in one direction and extraordinarily difficult to solve in reverse. But nobody has proved that the mathematical assumptions underlying the systems we use today must remain secure. We have simply failed, so far, to find sufficiently efficient ways of defeating them.

It is possible that somewhere in the universe of mathematical knowledge, an undiscovered algorithm exists that would make much of our digital security obsolete. An AI system might find it.

This is no longer an entirely abstract concern. In July, Anthropic announced that its Claude Mythos Preview model had helped discover improved attacks against two cryptographic constructions. One substantially weakened HAWK, a proposed post-quantum digital signature scheme. Another found a faster way to attack a reduced-round version of the Advanced Encryption Standard, the symmetric cipher that protects enormous quantities of data.

Neither discovery broke encryption used in production systems. HAWK was still a candidate under evaluation, while the AES result concerned a deliberately weakened version of an established algorithm. Anthropic also emphasized that the discoveries did not undermine other post-quantum signature candidates or lattice-based cryptography generally. Indeed, finding vulnerabilities before algorithms are deployed is one of the principal purposes of cryptographic research.

Still, something remarkable happened. A machine made original contributions to the mathematics of breaking codes. It would be rather complacent to assume that this line of inquiry will stop at the reassuring examples.

Consider what actually happens when you visit your bank’s website. Your computer and the bank’s servers must establish a shared secret, despite communicating through infrastructure that neither party controls. Public-key cryptography makes this possible. It allows two strangers to establish a secure channel without first meeting in a dark alley to exchange a password. Once that shared secret is established, symmetric encryption protects the information being transmitted.

Our commercial civilization depends on this arrangement. Banking, commerce, corporate communications, and secure messaging all rely on related cryptographic mechanisms. The extraordinary thing about the system is that it works without requiring a central authority to possess the secrets being exchanged.

This is one of the few genuinely decentralizing achievements of the digital revolution. A civilization otherwise increasingly dependent on enormous corporations and bureaucracies discovered a way to make at least some communications private by mathematical necessity.

Pimpale draws on a framework developed by complexity theorist Russell Impagliazzo, who imagined several possible mathematical worlds. In one, called Cryptomania, the mathematical structures necessary for public-key cryptography really do exist. In another, Minicrypt, certain forms of cryptography remain possible, but establishing a shared secret over an entirely public channel, without prior arrangements or trusted assistance, is not.

If we inhabit the latter world, the consequences would extend far beyond having to change our passwords.

Ordinary symmetric encryption would survive. People who had securely exchanged keys in advance could continue communicating privately. But the familiar ability to establish a secure connection with a complete stranger over the internet, without previously sharing a secret or relying on a trusted intermediary, would disappear.

Two people might exchange keys in person, perhaps by scanning a code on each other’s phones. This would be inconvenient for messaging and nearly impossible as a universal basis for online commerce. Alternatively, people could rely on intermediaries to distribute cryptographic keys on their behalf.

Such intermediaries would provide an obvious practical solution, although their precise design would matter enormously. A poorly designed system could give its operators access to the communications they facilitate. A more carefully constructed one might distribute trust among several independent organizations, requiring multiple parties to cooperate before sensitive information could be recovered.

There would be no mathematical requirement for a single universal gatekeeper. But the incentives to create one would be considerable, and governments have rarely been indifferent to the prospect of becoming indispensable.

Here we arrive at the political problem.

For decades, Western governments have been engaged in a mostly unsuccessful campaign to persuade technology companies that private communications should contain some provision for official inspection. The demands are invariably presented as reasonable. Nobody wants terrorists organizing atrocities, criminals trafficking children, or organized crime operating beyond the reach of law enforcement. These are real concerns, and governments have a legitimate responsibility to investigate them.

But there is a difference between obtaining a warrant to search a particular suspect’s property and requiring that every citizen’s communications infrastructure contain a mechanism for possible government access.

In 1993, the Clinton administration proposed the Clipper chip, a government-backed encryption technology designed to permit authorized surveillance through a system of escrowed keys. According to NIST’s own institutional history, the proposal met fierce public resistance. Americans were being asked to accept a communications system in which government agencies could recover the keys needed to decipher supposedly private conversations.

The Clipper chip failed to achieve widespread adoption, but the underlying ambition survived.

In 2017, Deputy Attorney General Rod Rosenstein argued that technology companies should retain the ability to provide readable communications and data when presented with an appropriate court order. He called this responsible encryption, insisting that security and lawful government access could coexist.

The argument has crossed party lines for decades. Surely the engineers who can make a message unreadable can also arrange for it to become readable when an appropriate official presents the necessary paperwork.

The trouble is that mathematics does not recognize the distinction between a benevolent government employee and a malicious intruder. A mechanism that allows one party to circumvent encryption creates capabilities that can potentially be stolen, abused, or exploited. A warrant may authorize access, but it cannot magically prevent an access mechanism from being compromised.

Britain offered a useful demonstration of the political conflict in 2025, when Apple announced it could no longer offer Advanced Data Protection to users in the United Kingdom. The decision came amid a dispute over British government demands for access to encrypted cloud data. Advanced Data Protection had extended end-to-end encryption to additional categories of iCloud information, including backups, photographs, and documents. Apple’s existing end-to-end encryption for services such as iMessage and FaceTime remained intact.

It is worth observing that the people losing this additional protection were not exclusively international terrorists. They included ordinary citizens storing family photographs, personal documents, and financial information. A safeguard intended for everyone became unavailable in Britain amid a dispute about the government’s ability to access particular people’s data.

Artificial intelligence could eventually transform this longstanding political conflict without a single parliamentary vote.

If public-key cryptography became unworkable, and ordinary internet users increasingly depended on intermediaries for secure communication, the pressure to regulate those intermediaries would be enormous. Governments could demand identification procedures, compliance requirements, and mechanisms for obtaining evidence. Large technology companies might find cooperation especially attractive if their privileged position depended on regulatory approval.

The result could be something remarkably close to the surveillance architecture governments have spent decades trying to legislate into existence, delivered instead by a mathematical discovery.

None of this would be inevitable. Distributed systems, independent key providers, and carefully designed protocols could prevent any single institution from controlling the process. But political institutions tend to favor arrangements that are convenient to administer, and centralized systems are generally more convenient for the administrator than for the administered.

There is an additional, rather unpleasant possibility. What if an intelligence agency discovers the breakthrough first?

The popular imagination tends to assume that scientific knowledge spreads outward, becoming available to everyone through the ordinary workings of research and publication. Intelligence agencies operate according to a different principle. A discovered vulnerability is useful precisely because the people exposed to it do not know it exists.

An undisclosed breakthrough against a widely deployed cryptographic system could offer a government an extraordinary advantage. It might continue allowing citizens, businesses, and foreign adversaries to believe their communications were secure while quietly exploiting the weakness. There is no public evidence that AI has already discovered a universal method of defeating public-key cryptography. But the incentive to conceal a consequential vulnerability, once discovered, is obvious.

And the danger is hardly confined to Washington. A mathematical vulnerability does not become an American vulnerability merely because an American company discovers it first. Chinese intelligence, Russian intelligence, criminal organizations, and independent researchers could all pursue the same problems.

This ought to complicate the increasingly fashionable assumption that the American government must have access to everything in order to keep Americans safe. The security of a nation depends on the ability of its citizens, businesses, military contractors, and public institutions to keep secrets from hostile powers. Building exceptional access into essential systems can undermine that objective, however admirable the immediate investigative purpose.

There are reasons for optimism. AI might help cryptographers discover weaknesses before hostile actors exploit them, and could eventually contribute to designing stronger cryptographic systems. Anthropic’s research is evidence of that possibility as much as it is a warning about offensive capabilities.

The National Institute of Standards and Technology has already finalized three post-quantum cryptographic standards, designed to address the threat that sufficiently powerful quantum computers would pose to existing public-key systems. In March 2025, NIST selected an additional algorithm, HQC, to serve as a backup to its principal standard for establishing shared encryption keys. The two rely on different mathematical foundations, reducing dependence on any single approach.

These are sensible precautions, although quantum-resistant does not mean immune to every possible future mathematical discovery. No collection of standards can guarantee that researchers, human or artificial, will never find a more efficient way to solve the problems on which those standards depend.

Pimpale suggests preparing infrastructure that could function without public-key cryptography, including arrangements in which several independent authorities would need to cooperate before access to protected communications became possible. There is merit in investigating these contingencies. A civilization that depends on cryptography should understand what it would do if its assumptions failed.

But preparing for technological failure must not become a pretext for building centralized surveillance infrastructure before that failure occurs.

The better approach is to preserve cryptographic diversity, invest in openly scrutinized research, develop systems that can replace vulnerable algorithms quickly, and ensure that any unavoidable intermediaries face strict limits on their authority. Congress should also examine how intelligence agencies would handle the discovery of a vulnerability threatening civilian communications on a national scale. There is a genuine national-security interest in secrecy, but there is also a national-security interest in preventing millions of Americans from relying on protections their own government knows have failed.

Conservatives, in particular, should resist the temptation to treat this as another dispute between law enforcement and Silicon Valley. There are good reasons to distrust the technology companies that have made themselves indispensable to modern life. There are equally good reasons to distrust a government that would prefer every private conversation to be recoverable upon request.

A constitutional republic cannot depend entirely on the benevolence of its administrators. Its institutions must be designed around the possibility that power will eventually be exercised by people less scrupulous than those presently holding it. The same principle ought to apply to the technologies through which modern citizens conduct their lives.

For all its faults, encryption has supplied something that neither corporations nor states have shown much enthusiasm for giving us: a private sphere beyond their immediate reach. It has done so through the peculiar authority of mathematics, which cannot be lobbied, intimidated, or persuaded to make an exception for national security.

That protection may prove more fragile than we imagined. If it disappears, we should be very careful about who offers to replace it. The state has always wanted to know what its citizens are saying. The terrifying possibility is that the machines might finally tell it.