Post-Mythos Crypto

Cryptocurrencies need a plan for a post-Mythos world.

Blockchains likely contain latent critical bugs, and AI systems are rapidly improving at finding them. At the same time, public blockchain infrastructure is unusually difficult to patch quickly or quietly.

Several recent evaluations from the UK AI Security Institute (AISI) on Anthropic’s latest model Mythos ⊕ Mythos is used here as shorthand for the first generation of frontier models that cross the threshold of reliable, autonomous, multi-step cyber capability. are showing AI systems are becoming substantially more effective at cyber capabilities: vulnerability discovery, exploit reasoning, patch analysis, and multi-step offensive workflows. To date, no major blockchain team has announced access to Mythos-like models.⊕ Please update the authors if you are aware of any.

Cryptocurrency teams are often international teams with no presence in the US, which might be incompatible with the current policies of some frontier labs.

Just as post-quantum safety demands a years-long, coordinated transition to new algorithms now, while we still can, we should address post-Mythos safety with the same urgency.

The thesis from this memo in short:

• AI with cyber capabilities is now gated to good actors, but soon will be more broadly available (e.g via jailbreak, via broader offer, or open source models catching up)
• Blockchains have latent bugs, and most should be fixed before cyber capabilities become broadly available
• Blockchain teams should prepare for a post-Mythos world where releasing bug fixes could become an attack vector (since blockchains are slow to upgrade and attackers can extract attacks from bug fixes).

The call to action for this memo is:

• Frontier labs should seriously consider giving cryptocurrency teams (or trusted security auditors) access to AI cyber capabilities during the grace period
• Security teams should put post-Mythos security at the top of their priorities (short term: finding bugs, long term: plan for secure upgrades)

Three periods after Mythos

In a pre-Mythos world, bugs are found by torchlight (expert security teams searching one at a time); in a post-Mythos world, they are found by radar (AI sweeping the whole surface at once).

Pre-Mythos is the world we used to live in. Finding a deep vulnerability in a mature system required a specialist with deep context — compiler behaviour, consensus internals, a specific cryptographic implementation. Very few people had that skill, audits were expensive, and that scarcity was, in effect, a security control. Bug density stayed bounded because the labour to exhaust it did not exist.

The grace period is where we are now. Mythos-class capability exists, but it is gated: it lives behind a small number of frontier labs that can choose who uses it and refuse misuse. In this window, defenders can get there first. For example: A protocol team with access can turn AI loose on its own consensus client, its proving system, its wallet stack — and fix most of what is cheaply findable before any attacker can run the same search. The grace period is precisely the interval during which access to Mythos-class capability can still be denied to bad actors. It is a temporary, expiring asset, and it is the central concept of this memo. Not having access to such capabilities in this phase will leave us unprepared for the next phase.

The unstable period begins when frontier access broadens enough, or open-weight models catch up enough, that anyone can run Mythos-class analysis. Now every latent bug in widely deployed software is in play simultaneously. This is the dangerous period, and it is not clear it will be survivable without severe exploited incidents. Two concerns dominate:

• Attackers find a bug before defenders do, and exploit it directly. At cryptocurrency scale, a single such bug in a shared component can mean irreversible, nine-figure loss.
• Defenders find and fix a bug first — and the fix itself leaks it. A good team patches quietly, but an attacker running Mythos-class patch-diff analysis reverse-engineers the vulnerability from the public upgrade, builds an exploit, and hits every node that has not yet upgraded before the fix propagates. For slow-to-coordinate blockchains, that propagation window can be days or weeks. (This is the secure-upgrade problem, below.)

The stable period is the post-Mythos equilibrium worth aiming for: a world where essentially every bug findable at reasonable cost by Mythos-class tooling has already been found and fixed, and where we have working answers to formally secure software and secure software upgrades. It is the analogue of a fully PQC-migrated internet. We do not get there by default; we get there by doing the work during the grace period.

What is actually at risk

It is worth being concrete about where in the stack the high-severity bugs live. These are the components where a single vulnerability is both hard to find by hand today and catastrophic if found by an adversary tomorrow:

• Consensus protocols. Consensus and execution clients, fork-choice logic, equivocation edge cases, p2p networking, mempool handling. A bug here can halt a chain, split it, or admit an invalid state transition.
• Complex cryptography. Proving systems and zk circuits, signature and commitment schemes, trusted-setup assumptions, custom or hand-rolled crypto libraries. The hardest code to audit and the easiest to get subtly wrong.
• Wallet infrastructure. Key generation and management, signing flows, browser-extension and hardware-wallet stacks. The bug that drains users directly, without touching the chain.
• Smart-contract code. The application logic itself — DeFi protocols, bridges, lending markets, DAOs — much of it immutable once deployed, publicly readable, and holding live funds. Reentrancy, access-control slips, arithmetic and rounding errors, and flawed economic assumptions are exactly the pattern-matchable bugs AI is good at surfacing at scale, and the high-value targets sit in the open for anyone to scan.
• Smart-contract compilers. Solidity, Vyper, Move compilers and their optimizers. A miscompilation silently breaks every contract built with it.
• Virtual-machine execution. EVM/SVM/MoveVM semantics, gas metering, precompiles. A bug here is inherited by everything that runs on top.

These are shared layers that large parts of the ecosystem depend on at once, and a low-level flaw propagates upward into every application that inherited its assumptions.

The secure-software-upgrade problem

Traditional software has an escape hatch that blockchains lack: it can patch quietly and fast. Vendors push automatic updates to millions of users, often before anyone notices; staged disclosure works because patch distribution outruns patch analysis.

Public blockchains invert both properties. Upgrades are slow — they may require coordination across validators, node operators, exchanges, RPC providers, wallets, bridges, and governance participants — and they are transparent, visible in public repositories and on-chain before they take effect. Historically that was tolerable, because only a handful of researchers could infer a vulnerability from a patch diff fast enough to matter. Mythos-class patch analysis destroys that assumption. The defender’s own fix becomes the attacker’s specification.

This is the crypto-specific crux, and it has no clean solution yet. We need upgrade mechanisms that are fast (propagate before the window closes), low-leakage (do not telegraph the vulnerability they fix), and coordinatable across decentralized, adversarial participants. That is a genuine, multi-year research-and-engineering migration — the direct analogue of rolling out post-quantum cryptography. It should start now.

Call to action

The grace period is a scarce resource that is being spent whether or not we use it well. Concretely:

1. Treat the grace period as expiring, and plan against its end. Decide now what your protocol does in the unstable period, before access broadens and the choice is made for you.
2. Get defenders Mythos access now. Frontier labs and policymakers should prioritize and fund vetted, gated access to Mythos-class cyber capability for protocol security teams — the people defending the highest-value shared infrastructure should be among the first, not the last, to wield it. This is the single highest-leverage action available during the grace period.
3. Build bug-finding and reporting protocols for an AI world. Continuous AI-driven auditing of the components above; disclosure pipelines designed for the reality that any published patch will be analyzed by Mythos-class tooling within hours.
4. Solve secure software upgrades. Fund the fast, low-leakage, coordinated upgrade mechanisms described above as a first-class research priority for the ecosystem.

None of this assumes cryptocurrency is doomed, however the trajectory is real and the grace period is short. Post-quantum cryptography is a decade of coordinated work done ahead of the threat. Post-Mythos crypto deserves the same seriousness, starting from the window we are in right now.

A stable post-Mythos world is possible: one where most bugs findable at reasonable cost by AI have already been found, fixed, or made unexploitable; and where software upgrades are themselves secure under adversarial scrutiny.

Get in touch

If you work on disclosure, upgrade mechanisms, or AI cyber-capability evaluations for blockchain ecosystems — or you want gated Mythos-class access for a protocol security team — reach out. DM @iamnotnicola on X.

Acknowledgements

This was written by Nicola Greco with support of AI.

Thanks for conversations to Alex Obadia, Jesse Walden, and Marco De Rossi.