Cyrus Reed: Rough week for post-quantum crypto, honestly — how are you holding up after reading this one?
Iris Holm: Concerned. Specifically about the framing.
Cyrus Reed: The hundred thousand dollar framing. Yeah. Okay so — Anthropic published this July 28th, their Frontier Red Team paper, and the core claim is: Claude Mythos Preview found a nontrivial automorphism in HAWK-256's lattice structure. HAWK being the post-quantum signature candidate currently under NIST review. And the attack effectively halves the key strength. Sixty hours. The same problem that beat two years of expert human review.
Iris Holm: Hold on — halves the key strength. Meaning HAWK-256 would need to double key sizes to maintain intended security?
Cyrus Reed: That's Anthropic's description, yes. And then they follow it immediately with — don't worry, not deployed yet, no production systems affected. And the cost barrier: a hundred grand per attack run.
Iris Holm: Which is reassurance theater if compute costs keep falling at the rate they have been.
Cyrus Reed: Okay but — I want to sit with the 60 hours thing for a second, because it's not just speed. The way Mythos did this was a multi-agent workflow. And in the AES part of the same paper, one worker agent literally rejected the key attack idea as infeasible, and a second one pursued it anyway and it worked. So the discovery came from — wait, is that the same thing as stochastic search? Or is something weirder happening?
Iris Holm: That's the second story inside this story. But the cost number is where I want to start, because it's doing a specific kind of work.
Cyrus Reed: But the cost number isn't actually the thing, is it — wait, you're about to tell me the thing.
Iris Holm: The thing is the category. The Frontier Red Team was explicit about it. Previous Claude demos found implementation bugs — a library with a crooked lock. Mythos attacked the math underneath the lock. The blueprint itself.
Cyrus Reed: Hold on — say that again without the jargon.
Iris Holm: Bank vault. Prior AI found the lock was installed crooked — fixable, patch it, move on. Mythos found the design of the lock is wrong. You can't patch a blueprint flaw. That's the distinction.
Cyrus Reed: HAWK-256, the attack doesn't say 'you misconfigured something.' It says the lattice structure has a nontrivial automorphism, meaning the underlying math hands you half the key for free. And then separately, on seven-round AES-128, the improvement isn't 'found a shortcut' — it's two hundred to eight hundred times better than the best known attack. That's not a bug. That's... wait, that's a proof that the structure was always slightly weaker than assumed?
Iris Holm: For that AES variant, yes. And the Frontier Red Team called both of these 'the strongest attacks we have found to date.' Their words. That's not hedged language.
Cyrus Reed: No way — they actually said that? Because that's Anthropic essentially grading their own find as a substantial research advance.
Iris Holm: Which is why the 'no production impact' framing frustrates me. Technically true — HAWK isn't deployed, the AES attack hits a reduced variant. But 'no production impact' and 'mathematically novel attack on a NIST candidate' are not the same claim dressed differently. One is a deployment status. The other is a capability signal.
Cyrus Reed: And HAWK has been in expert human review for two years — two years — and Mythos found this in sixty hours. That gap is actually the story, not the dollar figure.
Iris Holm: That gap is the story — but it's also where the stochastic problem bites back. Because here's what we don't know: if Anthropic ran the same workflow again tomorrow, does it find the same automorphism?
Cyrus Reed: Wait — you mean it might not? Like, the same model weights, same prompt, and it just... misses it?
Iris Holm: The AES worker pair is the evidence. One agent said infeasible. Another pursued it anyway. That's not a robust signal — that's one billion tokens of search space, and the answer happened to fall out. Florian Tramèr flagged this: longer than any standard benchmark evaluation, over a hundred thousand dollars. That's not a capability. That's a lottery with good odds.
Cyrus Reed: Okay but — wait, actually that framing might be too harsh. Because imagine a grad student locked in a room for two years with the HAWK spec. They also might reject a promising angle on Monday and pick it back up Thursday. The stochasticity isn't — I mean, is the stochasticity disqualifying, or is it just... how exploratory math works?
Iris Holm: Fair. But the grad student has one trajectory. The multi-agent workflow is running divergent trajectories in parallel. That's the actual win — not that it's smarter, but that it can explore branches simultaneously that a human would serialize.
Cyrus Reed: And one of those branches found a nontrivial automorphism that two years of expert human review serialized right past. That's — okay, that's the partial win I'll take. The result is real. The structure of how it got there is contingent, but the finding isn't.
Iris Holm: The finding is real. The capability is not yet proven repeatable. Those are different claims.
Cyrus Reed: No, that's right — and that distinction is going to matter a lot more when we get to why Anthropic published the finding but didn't release Mythos itself, because the asymmetry there is doing something structurally strange.
Iris Holm: One contingent discovery, disclosed without the tool that made it. That gap has a name.
Cyrus Reed: The gap has a name and the name is — wait, I want to say 'moral hazard' but I'm not sure that's exactly right. Because Anthropic publishes the HAWK finding, withholds Claude Mythos Preview entirely, and then gets credited for responsible disclosure. That's the asymmetry. They capture the reputational upside of transparency without any of the friction of actual scrutiny.
Iris Holm: And Robert Campbell saw it coming. January 2026 — months before the July 28th announcement — he published that Mythos-class models would compress PQC migration timelines. Cloud, enterprise, national security. The research community had already predicted the inflection point before Anthropic confirmed it.
Cyrus Reed: No way — Campbell's paper predates the announcement by six months?
Iris Holm: Which means Anthropic isn't alerting anyone to a new risk. The academic community was already there. Anthropic is confirming it — and doing so in a way that makes them look like the responsible actor.
Cyrus Reed: Okay but — here's what that means for NIST. FIPS 203, 204, 205 are already finalized. NSA CNSA 2.0 is an active federal mandate. The migration is running. And HAWK is still a candidate, not yet standardized. So the 'no production impact' claim is technically correct *today* — but if compute drops from a hundred thousand to, say, ten thousand before NIST finishes the migration, that qualifier just... expires.
Iris Holm: That's the defensible claim. Not 'AI broke crypto.' Not 'nothing to see here.' The 'no production impact' framing is a timeline claim, not a technical one. It holds only if NIST's migration completes before costs collapse.
Cyrus Reed: And Campbell's whole point was that Mythos-class capability *accelerates* the pressure on that timeline. The migration was already racing compute costs — now it's racing compute costs while someone just demonstrated the attack surface on HAWK is bigger than assumed.
Iris Holm: So the sting is this: Anthropic publishes the finding, keeps the model restricted, and frames it as safety-first. Campbell's January paper means that framing is six months late to be novel. And 'strongest attacks we have found to date' — their own words — is not reassurance. That's a capability floor, not a ceiling.
Cyrus Reed: A capability floor that costs a hundred thousand dollars today. And maybe ten thousand in eighteen months. That's — yeah, that's the actual clock.
Iris Holm: The clock is the point. Not whether it breaks crypto today — whether restricted access to Claude Mythos Preview outlasts the cost curve. That's the only bet Anthropic is actually making.
Cyrus Reed: And it's a bet they haven't — wait, they haven't named the terms of. Like, what's the condition under which they release Mythos? There's no stated trigger. It's just... withheld, indefinitely, while the NIST migration runs its timeline and compute costs do what compute costs do. That's a lot of load on an unstated assumption.
Iris Holm: You opened this asking how I was holding up after reading it. Honestly — rougher week than I expected. Not because the HAWK finding is catastrophic today. Because 'no production impact today' is a sentence that has an expiration date printed on it in invisible ink.
Cyrus Reed: Yeah. The invisible ink is the cost curve and the migration timeline, and neither one is Anthropic's to control. That's the crux of it. The responsible disclosure worked in July 2026. Ask me again when autonomous cryptanalysis is a ten-thousand-dollar compute job.
Iris Holm: Worth watching. Good conversation.
