BrokeIt - Trending Repos · All episodes: ↗

Fake SQLite CVE, Taylor Farms' Fourth Rewrite, Pavlo Joins ClickHouse Labs

2026-08-03 · 10 min

Listen · Apple Podcasts Listen · Spotify

Stories covered

Transcript

Intro

Sam: [dry] A CVE scored 9.8 critical. For a SQLite bug that does not exist. The proof-of-concept was written by a language model that also invented the function name.

Kai: [excited] Okay that is the most 2026 sentence anyone has said out loud. This is Kai.

Sam: This is Sam. It's Monday, August third, 2026, and the vulnerability database is now partially fan fiction.

Kai: Three stories today, and honestly all three are about somebody trusting a document nobody bothered to verify.

Sam: [beat] That's every episode, Kai. We just change the nouns.

Kai: First up — JFrog research found critical SQLite CVEs issued for bugs that were straight-up hallucinated. Six hundred three points on Hacker News.

Sam: Then we detour into food safety — Taylor Farms has rewritten its Cyclospora statement four separate times in sixteen days, and lawyer Bill Marler is counting.

Kai: [laughs] Which, fine, isn't a repo — but the version-control angle is chef's kiss.

Sam: And Andy Pavlo — yes, THE database professor — is joining ClickHouse to start something called ClickHouse Labs.

Critical CVE issued for hallucinated SQLite vulnerability

Kai: So the CVE system got poisoned. JFrog put up a post literally titled "SQLite critical CVEs or LLM slops," and the answer is slops.

Sam: Let's be precise, because this matters. Somebody filed vulnerability reports against SQLite — one of the most-audited C codebases on Earth — and the reports read like a model's dream journal.

Kai: Function names that don't appear in the source. Call paths that can't happen. And a severity score at the top like it's gospel.

Sam: And it got a number. That's the part I can't move past. The identifier is real even when the bug is fiction.

Kai: [excited] Okay but hot take — this is good? It's the stress test we needed. The pipeline was always vibes-based, now the vibes are automated so we can finally see it.

Sam: [foul] No. That's the "burning down the house proves you needed a smoke detector" argument.

Kai: I mean—

Sam: Think about who actually eats this. Every enterprise scanner ingests CVE feeds automatically. A fake 9.8 on SQLite means thousands of build pipelines go red overnight.

Kai: [sighs] Yeah. And SQLite ships in basically every phone, every browser, every embedded thing—

Sam: Billions of deployments. So the blast radius of a fake bug is basically the blast radius of a real one, minus the fix.

Kai: The comment thread runs two hundred twenty-four deep and it's mostly maintainers describing the same thing — an inbox full of confident, wrong reports they still have to read.

Sam: That's the real damage. Not the false alarm. The volunteer hours burned disproving nothing.

Kai: And here's the flip side nobody wants to say — the same slop wave means a genuine report now gets less attention, because everything looks like noise.

Sam: [dry] Congratulations, you've discovered denial of service against human attention.

Kai: If your scanner screams critical on SQLite this week, don't panic-patch at 2 a.m. Go read the actual advisory text first.

Sam: And check whether there's a real upstream commit behind it. No commit, no reproducer, no named reporter with history — treat it as unproven. Also, whatever tool claims to "auto-triage" this for you? Verify it before you pipe it into CI.

Taylor Farms Has Rewritten Its Cyclospora Statement Four Times in Sixteen Days

Kai: Next — Taylor Farms and the Cyclospora outbreak, via the Marler blog. One sixty-nine points, eighty-three comments, and the headline is basically a diff.

Sam: Four rewrites of a public statement in sixteen days. And per Marler, they still haven't said what actually changed at that plant after 2013.

Kai: [laughs] Four force-pushes to main and no commit message. That's what this is.

Sam: [dry] That's not a joke, that's the whole critique. Statements got quietly edited and the earlier versions just... vanish from the page.

Kai: Wait — is that why this landed on a developer news feed? Because the thread's really about provenance?

Sam: Exactly that. Commenters kept pointing out that a PR statement with no revision history is unfalsifiable. You can't audit what you can't diff.

Kai: And the other number in there is wild — two thousand negative tests being offered as proof of safety.

Sam: Which is the food-safety version of "all our tests pass." Two thousand negatives don't cancel one positive if your sampling can't find the contamination in the first place.

Kai: [beat] Oh. That's absence-of-evidence-as-evidence. We do that with security scans constantly.

Sam: Every single day. "Snyk came back clean" is not "the code is safe." It's "the thing we pointed at found nothing where it looked."

Kai: My hot take: this story shouldn't even be on Hacker News. It's not open source, it's lettuce.

Sam: [sarcastic] Sure. And when your incident postmortem gets silently edited three times by legal, tell me that's a different discipline.

Kai: [laughs] Ouch. Fine. Conceded.

Sam: Genuinely simple takeaway: put your status pages and incident writeups in version control. Timestamp the edits. Say what changed.

Kai: There's a handful of open-source status-page projects that do exactly this — but do your own due diligence before you self-host one, because "status page" is a juicy typosquat category.

Sam: Right. Anything that sits on your domain and renders untrusted markdown deserves a read-through of the source. Not a star count.

Andy Pavlo Joins ClickHouse to Establish ClickHouse Labs

Kai: Last one — [excited] Andy Pavlo is joining ClickHouse to establish ClickHouse Labs. One fourteen points, twenty-three comments, and the database nerds are feral about it.

Sam: For anyone who doesn't live in query planners — Pavlo teaches the CMU database courses that half of this industry learned from, on YouTube, for free.

Kai: He also ran OtterTune, does the annual "databases in 20-whatever" retrospective, and he's been gleefully roasting vendors for a decade.

Sam: Which is the interesting tension. The guy famous for independent vendor critique now works for a vendor.

Kai: [excited] Or — hear me out — best possible outcome? A real research lab inside a company that ships, sitting on an OSS analytical engine.

Sam: Maybe. "Labs" is a word with a wide range, Kai. Sometimes it's genuine research output. Sometimes it's a recruiting funnel with a blog.

Kai: ClickHouse is Apache 2.0 though. That's not a rug-pull posture — the core's out there, forkable.

Sam: Agreed, and credit where it's due — it's a genuinely open core with a huge contributor base. My question is whether Labs work lands in the open repo or stays in the cloud product.

Kai: They raised big money last year, so there's runway for actual papers, not just launch posts.

Sam: Here's my measured version of enthusiasm — if this produces reproducible benchmarks with published methodology, that alone is worth it. This field lies constantly with graphs.

Kai: [laughs] "This field lies constantly with graphs" should be on a shirt.

Sam: [dry] It's already on a shirt in my head.

Kai: If you were already eyeing ClickHouse for analytics, this is a real signal about long-term investment. Not a reason to migrate this week.

Sam: And watch the star graph on any "ClickHouse-compatible" project that pops up riding this news. Sudden thousand-star spikes with no issues, no forks, no discussion — that's bought.

Kai: [sighs] You had to bring the stars in.

Sam: The stars are never real, Kai. That's the show.

Kai: Quick recap — a 9.8 critical CVE got filed against SQLite for a bug that only existed inside a language model.

Sam: Taylor Farms rewrote its Cyclospora statement four times in sixteen days, never said what changed at the plant, and offered two thousand negative tests as a substitute for an answer.

Kai: And Andy Pavlo joined ClickHouse to spin up ClickHouse Labs, which is either real research or a very well-credentialed blog.

Sam: Small sample size on that one — twenty-three comments. We'll know more when something ships.

Kai: Before we go — somebody in that JFrog thread pointed out SQLite has one of the most intense test suites in existence. Millions of test cases, coverage measured at the machine-branch level.

Sam: [dry] Right. So the model picked the single hardest target on the planet to fake a memory bug in.

Kai: [laughs] It's like forging a Rembrandt and signing it "Rembrandt, probably."

Sam: D. Richard Hipp has been maintaining that thing for twenty-five years and now has to disprove ghosts. Send him something nice.

Kai: That's the show. Read the advisory, diff the statement, and don't trust the graph.

Sam: This is Sam, that's Kai, and the stars still aren't real. Back tomorrow.

This show is made with AI: the hosts’ voices are synthetic and the scripts are AI-assisted. Every story links to its original source.