See pricing
Can AI Break Cryptography? What the Experts Actually Think
The a16z ShowSecurity

Can AI Break Cryptography? What the Experts Actually Think

AI's gains in mathematical reasoning have triggered fresh fear that encryption could fail sooner than expected. Dan Boneh, Eddy Lazzarin, and Justin Thaler explain why that fear is still speculation, what evidence would actually matter, and why adaptable systems beat panic reactions.

October 13, 2026 · 42 min listen · 13 min read · Dan Boneh, Eddy Lazzarin, Justin Thaler
0:00
–:––

Context

This episode examines a question that sounds existential for any digital product team: could rapidly improving AI break the cryptography behind banking, messaging, cloud security, and crypto wallets? Stanford cryptographer Dan Boneh joins a16z crypto's Eddy Lazzarin and Justin Thaler to separate actual evidence from fear triggered by recent AI progress in mathematics. It matters to PMs because the discussion is really about risk management under uncertainty: how to design systems that depend on cryptographic assumptions, how to communicate about scary but unproven threats, and how to build products that can adapt if one security assumption fails.

The Big Idea

AI has not broken modern cryptography. What AI changes is the speed at which the world can test cryptographic assumptions, so the right response is better evidence and stronger crypto agility, not emergency moves based on speculation.

The panel's core point is that recent AI results did not show a direct attack on deployed cryptography. The more important shift is that frontier labs may now be able to validate or challenge security assumptions much faster than the old process of waiting years for the research community to probe them.

Key Insights

No break was announced

  • The recent concern came from AI progress in hard math and computer science, not from a published result that broke elliptic curves, RSA, lattices, or another deployed cryptographic scheme.
  • The speakers stressed that OpenAI's release showed broad technical progress, but it did not include a direct cryptography result that would justify panic.
  • For a PM, this distinction matters. A team should treat the moment as a change in the threat-monitoring environment, not as a confirmed security event.

Cryptography rests on assumptions

Modern cryptography is built on mathematical hardness assumptions. Teams assume problems like elliptic curve discrete log or lattice problems are hard because nobody has found an efficient way to solve them, not because hardness has been formally proved. The reason that proof is missing is tied to deep open questions like P versus NP. That means every product built on cryptography is, in some sense, built on a stack of strong but unprovable assumptions.

AI changes validation speed

  • Historically, confidence in a new cryptographic assumption came slowly. Researchers would attack it for years, and if nobody broke it, confidence rose over time.
  • The panel's argument is that AI could compress that process dramatically. A frontier lab could aim a top model at a cryptographic problem for weeks or a month and produce much faster evidence.
  • That evidence cuts both ways. If AI finds a new attack, teams need to react faster. If AI fails after serious effort, teams may gain more confidence than they would have gotten from a decade of ordinary scrutiny.

A real break is global

  • If elliptic curve cryptography were materially broken, the damage would extend far beyond crypto wallets.
  • Dan Boneh explicitly said the blast radius would include banking, healthcare, government communications, and the broader internet.
  • That framing is important for PMs. A cryptography failure is usually not a single-product incident. It is an ecosystem incident that hits vendors, protocols, infrastructure, and customer trust all at once.

Existing migrations already help

A useful calming point in the discussion is that the world is not starting from zero. NIST is already standardizing alternative public-key systems, and some deployed systems already use hybrid protection. The panel mentioned internet handshakes that combine elliptic curves with lattice-based methods, and pointed to products like iMessage and Signal as examples of migration already underway. In practice, that means many teams are already building the upgrade paths they would need if one assumption weakens.

AI is attacker and auditor

  • AI may help search for attacks on current schemes, including some post-quantum candidates.
  • The same AI can also improve defense by scanning software and cryptographic implementations before release, after code changes, and again when stronger models appear.
  • The panel described companies already rescanning codebases when software changes and when a new model becomes available. The long-term bet is that software quality and cryptographic quality should improve if teams use AI aggressively on defense.

Mental Models & Frameworks

Security by assumption

Use this model whenever a team talks about cryptography as if it were absolute.

  • What it means: cryptographic security usually rests on assumptions that certain problems are computationally hard.
  • Why it matters: a product team cannot eliminate that uncertainty, only manage it through careful scheme choice, migration paths, and monitoring.
  • How to use it: whenever a design depends on a signature scheme, key exchange method, or proof system, ask which exact hardness assumption the design inherits and what happens if that assumption weakens.

Confidence from failed attacks

This is the panel's main way of thinking about why cryptography becomes trustworthy over time.

  • Researchers gain confidence when many serious attack attempts fail.
  • AI can make that process faster by increasing the amount of scrutiny a scheme receives in a short time.
  • This flips the usual story. Better attackers are not only a threat. Better attackers also produce better evidence about what survives attack.

Crypto agility

Crypto agility means designing systems so they can switch from one cryptographic mechanism to another without rebuilding the whole product. Dan Boneh described it as a long-standing defensive principle and the right answer to unexpected breaks. Use it when choosing libraries, protocols, wallet architecture, authentication systems, or messaging infrastructure. If a product cannot rotate away from a broken primitive quickly, the real weakness is often architectural, not mathematical.

Hybrid migration path

A hybrid migration path uses more than one protection method during a transition period.

  • How it works: a connection or system can rely on both an existing scheme and a newer alternative, so breaking one does not immediately expose the system.
  • Why it matters: hybrid deployment buys time. Teams do not need perfect certainty before beginning a migration.
  • Example from the episode: the panel mentioned deployed internet connections that use elliptic curves and lattice-based cryptography together.

Decision Principles

Principle: Separate evidence from fear

  • When: a new AI result or security thread produces urgent public speculation.
  • Why: indirect evidence of AI progress is not the same as evidence that a cryptographic assumption has failed. Product teams should increase monitoring and information gathering, but should not communicate as if a breach has already happened.

Principle: Design for assumption failure

  • When: a product depends on a specific signature scheme, key exchange method, wallet design, or proof system.
  • Why: no current public-key assumption comes with a final proof of hardness. The safest product posture is to assume that some assumptions may eventually weaken and to build upgrade paths in advance.

Principle: Use AI defensively first

  • When: shipping security-sensitive code, cryptographic integrations, or protocol changes.
  • Why: the panel's strongest practical optimism is that AI can raise the baseline quality of software and cryptographic implementations. Defensive use is available now, even while the offensive ceiling remains uncertain.

Trade-offs & Nuance

Hashes feel safer, do less

  • Hash-based signatures appeal to many blockchain teams because blockchains already rely on hash functions in other places, such as Merkle-tree style structures.
  • If a specific hash function weakens, teams may be able to swap in another while preserving the broader framework.
  • The downside is capability. Hash-based systems are often larger, harder to use, and provide fewer algebraic properties for advanced constructions.

Lattices enable richer systems

Dan Boneh argued that lattice-based systems are attractive because they support more advanced features. He specifically pointed to areas like threshold signatures and key-derivation related structures. The trade-off is that if the underlying lattice assumptions are weaker than expected, a much larger family of constructions can fall together. So the choice is not just "more secure" versus "less secure." It is also simplicity versus expressive power.

Simple assumptions, complex machinery

Justin Thaler added an important caution to the "hashes only" instinct. In some blockchain designs, relying only on hashes pushes teams toward more complex machinery such as SNARK recursion. That may reduce one category of mathematical assumption while increasing implementation complexity, bug surface, and uncertainty about the security of the full construction. For a PM, the lesson is clear: reducing one theoretical risk can increase a very practical engineering risk.

Common Mistakes

Mistake: Confusing adjacent progress

The biggest error in the public reaction was treating broad AI progress in mathematics as if it were direct evidence that deployed cryptography had already become vulnerable. The panel's view was much narrower. AI progress changes what is worth investigating and what information the labs should share. It does not, by itself, prove that any particular scheme is broken.

Mistake: Pushing risky user moves

  • The speakers argued that ordinary users should not start making ad hoc migrations because of speculation alone.
  • Eddy Lazzarin noted that even Vitalik had pointed out a practical truth: people often lose funds in botched migrations more often than in the tail-risk event they are trying to avoid.
  • Better hygiene for institutions is useful. Asking normal users to do something risky without concrete evidence can create a larger immediate danger than the hypothetical future threat.

Practical Application

Map cryptographic dependencies

Create a living inventory of where the product depends on specific cryptographic assumptions.

  • List signature schemes, key exchange methods, wallet designs, proof systems, secure messaging dependencies, and vendor libraries.
  • Note which user journeys depend on each one, such as login, transaction signing, API auth, or encrypted sync.
  • This turns a vague panic topic into a concrete dependency map your team can act on.

Add a crypto switch plan

For each critical cryptographic dependency, document the replacement path before a crisis arrives.

  • Identify whether the dependency can move to a hybrid mode first.
  • Record what code, vendor, customer workflow, and legal or compliance changes a migration would require.
  • Set clear triggers for action, such as a standards update, a vendor announcement, or a credible published attack.

Turn AI on reviews

Use the strongest available models to review security-sensitive changes as part of the release process.

  • Run scans before deployment on the code most likely to create irreversible loss, such as signing flows, key management, authentication, and cryptographic wrappers.
  • Re-run scans after major code changes and when a meaningfully stronger model becomes available.
  • Treat model output as a force multiplier for human review, not as a replacement for it.

Separate consumer and institution guidance

Write two levels of security response playbooks.

  • A consumer-facing playbook should explain what changed, what did not change, and whether any action is actually required.
  • An institution-facing playbook can include stricter hygiene, such as minimizing unnecessary public-key exposure where that applies.
  • This prevents a common failure mode where expert-only precautions leak into messaging for users who would be harmed by taking them.

Review public key exposure

If the product touches wallets or blockchain security, review where public keys become exposed and whether that exposure is necessary. The panel described limiting exposed public keys as good hygiene that many serious institutions already follow. Even if AI changes nothing in the short term, reducing unnecessary exposure is still a sound defensive step.

Questions to Consider

  • If elliptic curve signatures weakened tomorrow, which specific customer flows in your product would fail first: login, wallet signing, device pairing, secure messaging, or infrastructure TLS dependencies?
  • Can your engineering team swap one cryptographic primitive for another without rewriting core business logic, or is security tightly welded into the product architecture?
  • Are you reducing one mathematical assumption by adding a much more complex construction, such as recursive proof machinery, that creates a larger implementation bug surface?
  • Where would AI-based code and cryptography scanning save the most security review time in your current release process?
  • If a frontier lab privately found a real attack, would your team know which vendor, protocol owner, or standards body to listen to first before communicating with customers?

Bottom Line

The episode's answer is calm but not complacent: AI has not broken cryptography, but it may compress the time between a hidden weakness and a discovered attack. The winning product posture is to build systems that can switch assumptions quickly, use AI to harden software now, and wait for real evidence before asking users to do anything risky.

Resources Mentioned

ResourceTypeWhy it was mentioned
Original a16z crypto discussion on YouTubeVideo discussionThe feed links to the longer discussion that this episode draws from.
NIST post-quantum cryptography standardizationStandards effortMentioned as the main backup path if current public-key assumptions weaken and systems need to migrate.

Concepts to Explore

P versus NP

P versus NP is the famous open problem that sits in the background of modern cryptography. The episode matters more if you understand why cryptographers cannot simply prove that their preferred hard problems are truly hard. That gap is why the field relies on assumptions plus years of failed attack attempts.

Post-quantum cryptography

Post-quantum cryptography refers to schemes designed to survive attacks from large quantum computers. The episode adds a new wrinkle: a scheme can be post-quantum and still not necessarily be post-AI. That makes post-quantum migration necessary, but not sufficient, as a long-term security strategy.

SNARK recursion

SNARK recursion is a technique where one proof verifies another proof, and so on. It comes up here because some blockchain communities use it to make hash-heavy designs practical. The concept is worth exploring because it shows how removing one assumption can force much more complicated engineering elsewhere.

People to Follow

Dan Boneh

Dan Boneh is a Stanford professor and one of the best-known cryptographers in the field. In this episode he provides the clearest technical framing of why the right response is calm, evidence-seeking, and focused on crypto agility.

Justin Thaler

Justin Thaler is a Research Partner at a16z crypto. He adds the most useful nuance on how AI changes the process of gaining confidence in cryptographic assumptions and why simpler-looking replacements can still hide major complexity risk.

Eddy Lazzarin

Eddy Lazzarin is a General Partner at a16z crypto. His contribution is especially useful for PMs because he keeps pulling the discussion back to operational reality: bug risk, user behavior, responsible disclosure, and how security teams should communicate under uncertainty.

Experiments to Try

Experiment: AI red-team one module

  • Hypothesis: model-assisted review will find meaningful issues in one security-critical component that standard review either missed or took much longer to find.
  • Change: pick one narrow target, such as transaction signing, key management, or authentication.
  • Test: run your normal review process, then run a separate AI-assisted red-team pass with explicit prompts about cryptographic misuse, unsafe defaults, and migration risks.
  • Measure: count unique findings, severity of findings, time spent by reviewers, and how many AI findings survive human validation.

Experiment: Assumption-failure tabletop

  • Hypothesis: a written crypto-agility plan will cut response time and confusion if one cryptographic primitive becomes questionable.
  • Change: run a one-hour tabletop on a scenario such as "credible attack published against elliptic curve signatures."
  • Test: ask product, engineering, security, and support to identify affected flows, replacement paths, customer messaging, and vendor dependencies.
  • Measure: time to produce a dependency list, number of blocked migrations, number of missing owners, and clarity of customer communication.
AI PM course

Everyone hears the same episodes.
Few can do what they describe.

Start for free