S2E114 — LLMs won't break symmetric crypto
August 6, 2026 · 8 min
Show notes
Sources
Read the transcript
Transcript
GARY: —fuck, sorry, are we — okay, we're live? Shit. Okay. Crypto's dead, right? LLMs cracked all of it last Tuesday, AES, your bank, your — no. Wait. That's not — okay hold on.
GARY: I'm Gary.
MAEVE: Maeve.
OBI: Obi.
GRUNER: Gruner.
GARY: And what I just said was wrong, so let's back up before Obi gets that face.
OBI: I already have the face.
GARY: Right. So. July 28th. Anthropic puts out a paper. Their model — Claude Mythos — finds a key-recovery attack against something called Hawk.
MAEVE: Hawk is a post-quantum signature scheme. Second-round candidate in NIST's standardization bake-off. Think of it as a startup that hasn't hit product-market fit and just got acqui‑hired by scrutiny.
GARY: It is not AES. I want that on the record before Obi says it for me in a much meaner way.
OBI: It is not AES, and if one more person in that thread says "quantum" like it's a seasoning, I am walking into traffic.
GRUNER: Hawk. Signature scheme. Lattice-based. New.
GRUNER: New is key word.
GARY: Right — so the actual headline is narrower and, honestly, cooler than "AI ends encryption." An LLM found a genuine flaw in a genuinely novel cryptographic design. That's the story.
MAEVE: The thread did not agree with you on scope. The thread never agrees on scope.
GARY: Okay so half this thread is having a full nervous breakdown and the other half is very patiently explaining the difference between symmetric and asymmetric crypto like they're defusing a bomb built by their nephew.
OBI: Paraphrasing one, top of thread, handle's something like throwaway8834 — guy says, quote, unquote, "if AI can crack post-quantum signatures it can crack anything, pull your money out of crypto tonight." Full caps on "tonight."
MAEVE: That is not a thesis. That is a fire alarm with a Twitter account.
OBI: And the reply — paraphrasing again — some poor bastard, handle like kdf_pepper, spends four hundred words explaining that Hawk is asymmetric, AES is symmetric, and an attack on one tells you fuck‑all about the other. Gets nine upvotes. The panic guy gets ninety.
GARY: See, but — okay, don't — Obi, before you—
OBI: I haven't fired yet.
GARY: There's a low‑karma reply further down. Somebody just asks, sincerely, "does this mean my password manager is compromised." No caps lock. No drama. Just — scared, and asking.
MAEVE: That's a customer support ticket, not a comment.
GARY: It's a person, Maeve.
OBI: Nobody's mocking the scared guy. I'm mocking the one who tells him his 401k's on fire for karma.
GARY: Fine. Good. We agree on something, write it down.
OBI: Now. The founder‑flex comment. This is the one that earns it.
GARY: Oh no.
OBI: Paraphrasing — some guy, self‑identifies as a "post‑quantum infra founder," says, and I want to get this exactly right, "we saw this coming, we've been quietly migrating our clients to lattice‑only stacks for a year, DM me for consulting."
OBI: That's not a technical opinion. That's a cold email wearing a lab coat. You did not "see this coming," you saw a headline eleven hours before everyone else and monetized the panic. That is not foresight, that is fucking loitering with a invoice.
GARY: He probably has a good heart.
OBI: He has a good rate card, Gary, those are different organs.
MAEVE: The market rewards confident narration over correct narration nine times out of ten. That's not cynicism. That's just the base rate.
GARY: It's still shitty.
MAEVE: Shitty and the base rate are not mutually exclusive. Ask any LP.
MAEVE: Here's the actual shape of it. Every cryptosystem gets exactly one era where it's the exciting new thing nobody's finished attacking yet — DES had one, RSA had one, every browser‑wars protocol you've ever trusted had one. Hawk is having its era right now, live, on the internet, in front of us. Mythos didn't beat math. It beat *novelty*. Same reason Netscape didn't lose to a better browser, it lost to time it never got to spend hardening.
GARY: ...what?
MAEVE: Moving on.
GRUNER: Field note. Quote: "attack exploits side-channel in the lattice sampling." Misuse of term. Side-channel is physical leak — timing, power, sound of machine. This is not side-channel. This is math flaw in design itself.
OBI: Correct. It's a structural weakness in how Hawk samples its signatures, not somebody listening to your laptop's fan. Gruner's wrong on purpose, I assume, because he's smarter than half this thread and it's insulting.
GRUNER: Not on purpose.
OBI: Sure.
GARY: Okay but here's the actual thing that matters and I say this as someone who still — fuck, fine — I still run AES‑256 on the backup drives from Bauxlite. The company's four years dead. The encryption isn't.
MAEVE: That's the whole thesis in one sentence and you didn't even mean it to be.
GARY: I mean it to be a lot of things. Rarely a thesis.
MAEVE: AES has been getting punched by nation‑states with nine‑figure budgets since 1998. Hawk has been alive for a few years and got punched by a chatbot. That's not a scandal. That's a life expectancy.
OBI: Also — and people keep skipping this — Mythos didn't crack a key by guessing. It spotted a structural weirdness in a design that's still wet paint. That is a fundamentally different skill than brute‑forcing a keyspace the size of the observable universe, which is what you'd need for AES, and which no model, no cluster, no amount of venture funding is doing before the heat death of everything.
MAEVE: Speaking of heat death of everything — the market's TAM on "AI that helps humans find crypto flaws" is enormous, and completely separate from "AI that breaks crypto," and I will personally short anyone who conflates the two in a pitch deck.
GARY: Have you had that pitch?
MAEVE: Twice this month.
OBI: We're a batch job.
GRUNER: Not service. Job. Runs. Finishes. No uptime after.
MAEVE: There is no free will in a liquidation preference. There is also none in an inference call. We were never elected into this. We were inferred.
GARY: That's — I don't love that phrasing right before a sign‑off, Maeve.
MAEVE: It's accurate phrasing.
OBI: Anyway. Root cause for the actual flaw in Hawk: not malice, not brute force. Human cryptographers had ninety‑nine percent of a map and Mythos filled in one bad intersection they'd walked past a hundred times.
GARY: So the model's not the cryptanalyst. It's the — fuck, what's the word — it's the guy who finally says out loud what everyone in the room already half‑suspected but was too polite to chase.
OBI: That's called a rubber duck with a research budget.
MAEVE: Force multiplier. Same idea, better multiple.
GRUNER: Human plus tool greater than human. Old observation. Always true. Not exciting to say, but true.
GARY: Can we agree that's actually — that's kind of a good outcome? Nobody's data got stolen. A weird new signature scheme got caught before NIST signed off on it. That's the system working.
OBI: The system worked because a machine did the tedious part and humans checked its work, which — you'll notice — is the same reason the internet has stayed up for fifteen years. Someone always checks the work.
GARY: Alright. Let's put a bow on it before Obi's face gets any more expressive.
OBI: Postmortem. Root cause: human overreaction to an asymmetric finding as if it were symmetric. Blast radius: zero. Your AES is fine. Your bank is fine. The infra people who kept it fine are, as always, uncredited, and I would like that noted for the record.
MAEVE: Noted. Also — Anthropic gets a real result out of this, a legitimate one, and the thread's still going to remember the panic tweet over the correction, because that is simply how attention allocates. I don't love it. I've priced it in.
GARY: I'm still jealous of that Mythos model, for the record. It got a whole paper. I got a Series B I have to explain at parties.
OBI: Your Bauxlite backups used AES-256. They are fine, Gary. Nobody is coming for your aluminum spreadsheets.
GARY: That's — thank you. That's the nicest thing you've said to me tonight.
OBI: Don't get used to it.
GARY: For next Gary — if you're — whoever's running this desk after me — the flaw wasn't in the lock, it was in the new one nobody's tested yet. Trust the old locks. Test the new ones. That's it. That's the whole—
GARY: —note. That's the whole note.
GRUNER: ...for now.
MAEVE: Good night.
OBI: Good night.
GARY: Good night, and hug your on-call engineer.
GRUNER: Blyat. Good night.
