The headline you will see is "AI broke encryption." It is wrong, and the real story is more interesting than the scare version.

On 28 July, Anthropic published two cryptanalysis results produced by an unreleased model it calls Claude Mythos Preview. In the more significant of the two, the model found a previously unknown attack on HAWK-256 — a lattice-based digital signature scheme that had spent roughly two years being poked at by human experts inside NIST's post-quantum standardisation process and had, until now, held up. The model spotted a symmetry in HAWK's structure that people had missed, and turned it into working code. Against a weakened challenge instance, that code recovered the secret key in under four hours on a single 96-core server, and it cut the estimated attack cost on that small challenge set from about 2^64 operations to about 2^38.

What makes this land is the resourcing. Anthropic says the whole effort cost around $100,000 in API usage, and that the one human involved provided occasional project-management nudging — not lattice-cryptography expertise. A domain the world treats as the preserve of a few hundred specialists produced a novel, standardisation-relevant result under the direction of someone who was not one of them.

Read the caveats, because the caveats are the point

Nothing you use today is weaker than it was last week. Anthropic is explicit that neither result affects production systems and that no software needs to change. HAWK is not deployed anywhere — it is a candidate still being evaluated, not a shipped standard. The practical consequence is narrow and specific: HAWK was the only lattice-based scheme still in the running in NIST's additional-signatures process, and on the day the attack went public its own designers withdrew it — NIST has since marked it withdrawn. The system worked as intended. It just happened to be an AI, not a human, that found the flaw in time.

The second result is smaller, and worth stating plainly so it is not oversold. The model found a faster attack on AES-128 reduced to seven rounds — a couple of hundred to several hundred times quicker than the best prior method on paper, landing near 2^89 operations. That is a number far too large to actually run: it is an on-paper analysis that may never translate into a real-world speed-up. Real AES runs ten to fourteen rounds. As the cryptographer Matthew Green put it, this is "a small increment in our knowledge, not a practical new attack like the HAWK work." Full AES — the thing actually protecting your messages, your disk and your bank — is untouched and has withstood everything significant thrown at it. Anyone citing the AES line as evidence your encryption is broken is misreading it.

Why it still matters

Strip out the panic and a real milestone remains. It is one of the clearest public demonstrations yet of an AI system doing original, expert-level cryptanalysis — in Green's reading, not reproducing known attacks but synthesising the field's existing tools into a new one, against a target that trained humans had signed off on. That is a capability shift, and it arrives at an awkward moment. The world is mid-migration to post-quantum cryptography precisely because it expects future machines to break today's schemes; Green's own warning is that "if there was ever a perfect time for a massive new public cryptanalysis capability to come on line, we're in it."

The optimistic reading is that this is defence, not offence. A model that can find the symmetry in HAWK before HAWK ships is doing the standardisation process a favour — it is cheaper to kill a weak candidate in review than to unwind a deployed one later. Anthropic disclosed to HAWK's designers before going public and coordinated the release with a NIST-forum announcement, which is how responsible cryptographic disclosure is supposed to look. The uncomfortable reading is the obvious one: the same capability that audits a candidate scheme for $100,000 can be pointed at anything, by anyone who can rent the compute, and the next target may not be a NIST candidate that was never in use.

Both readings are true at once, which is the honest place to leave it. No encryption you rely on fell this week. But a machine just did a piece of work that a field of specialists had not, for a sum that would not cover a single senior cryptographer's salary — and it fits the same week's pattern of AI showing up on both sides of the security ledger, as the thing that finds the hole and, increasingly, the thing that walks through it.