#134

GLM-5.3 pulled 2,436 bugs out of 269 projects, and Firefox keeps uBlock Origin's one API

Z.ai partners logged 2,436 vulnerabilities in 269 projects, 2,383 still embargoed. Firefox keeps blockingWebRequest. Google's HEIR costs 24 seconds per inference.

Listen to this edition

Z.ai and its partner security teams pointed GLM at 269 open source projects and came back with 2,436 vulnerabilities. Only 53 of them are public. The rest sit under embargo, on flaws that had been in the code 26.6 years on average.

GLM-5.3 runs on the same base model as GLM-5.2 a month earlier, and every gain came from post-training. The weights ship two weeks after launch, once safety evaluation and hardening are done.

In today’s indie hacker news:

  • 🐛 GLM logged 2,436 bugs, 53 of them public
  • 🧱 Firefox is the last browser running full uBlock Origin
  • 🌑 A cryptographer says safer software is the problem
  • 🔐 Google’s compiler bills private inference in seconds
  • 📚 Mark Dominus keeps the thinner thesaurus

TOP STORIES

🐛 THE CALL IS COMING FROM 1981

A tall stack of sealed vulnerability reports with one dusty page glowing near the bottom

Z.ai says its models found thousands of real bugs, and published almost none of them.

The story: Z.ai’s launch post buries the interesting part under the benchmark tables. Since GLM-5.2, Z.ai has worked with security teams in China. Every finding goes into a public disclosure ledger at cvd.z.ai. The severity split runs 107 critical, 990 high, 1,286 medium. The oldest flaw dates to 1981, which the ledger card counts as 45 years of impact.

Semgrep ran it the same day on its IDOR detection benchmark. That’s the bug where changing an ID in a URL hands you someone else’s record. It scored 23.8% F1 against Claude Opus 4.8’s 23.6%, at $0.15 per confirmed true positive against $1.04. Same detection tier, roughly a seventh of the price.

The details:

  • Recall is the wall, not precision. GLM-5.3 posts 81.6% precision against 13.9% recall on that benchmark. It’s usually right when it flags something, and blind to most of the rest.
  • The independent run went backwards. Semgrep’s score landed below GLM-5.2’s 26.8% F1, the opposite direction from Z.ai’s own claims. Semgrep says it hasn’t confirmed whether that’s a real step back or run to run noise, and is re-running both.
  • The frontier is still far ahead. On the same leaderboard, Claude Opus 5 posts 65.6% F1 at $0.44 per true positive. Cheap detection and good detection are not the same purchase.
  • Off-peak is half price. The GLM Coding Plan bills points, and anything outside 14:00 to 18:00 UTC+8 on weekdays costs half. Overnight scans are the cheap slot.
  • Nobody has audited the ledger. Only the disclosed findings can be checked, so the total rests on Z.ai’s own count.

“Both fable 5 and opus 5 have outright refused things like security related bug fixes and making monitoring tools. I am so happy that open models are good now” (takerofnaps, Hacker News commenter)

Why builders care: Security review of your own repo is now a background job, not an enterprise line item. The same price works for whoever points it at what you shipped, starting the day the weights land.

🧱 LAST BLOCKER STANDING

A row of browser extension icons going gray with one orange puzzle piece still lit

Google pulls every remaining Manifest V2 extension on August 31, and Firefox keeps the API uBlock Origin runs on.

The story: Google’s own timeline removes every remaining Manifest V2 extension from the Chrome Web Store on August 31. Chrome stopped running MV2 a year ago, so this is the cleanup step. Microsoft announced on August 7 that Edge starts switching MV2 off for consumers this month. Consumer completion is targeted for the end of 2026, enterprise for early 2027.

That leaves Firefox. uBlock Origin runs on blockingWebRequest, and Chromium’s replacement, declarativeNetRequest, limits how an extension can filter content. Mozilla committed in February 2025 to supporting both. PCWorld’s read: Firefox is the only major browser left where uBlock Origin still works. Safari and DuckDuckGo never supported it.

The details:

  • Microsoft counted the survivors. Only 58 MV2 extensions on Edge Add-ons still have meaningful usage, and 3 of those lack a public MV3 build. Microsoft doesn’t name the three.
  • 95 percent already moved. That’s Microsoft’s stated justification: the top MV2 extensions in its store shipped MV3 versions before the cutover started.
  • Chrome closed the door in 2025. Chrome 138 was the last version that could run MV2. Chrome 139 removed the enterprise policy that kept it alive.
  • The Firefox-only surface is specific. webRequest.filterResponseData, CNAME uncloaking, and filtering by DNS record all live in Firefox alone, per uBlock Origin’s own wiki. CNAME uncloaking unmasks trackers disguised as first-party servers.
  • Mozilla’s stake is its install base. Nearly half of Firefox users have at least one extension installed.

“Our support for uBlock Origin isn’t going anywhere.” (Firefox, on Bluesky)

Why builders care: If you sell a blocker, it’s now a Firefox product plus a Lite build everywhere else. That makes addons.mozilla.org your primary install funnel, and one vendor’s blog post your only guarantee.

🌑 TOO SECURE TO TAP

A vault door sealing shut while a surveillance antenna behind it goes dark

A Johns Hopkins cryptographer says AI will make software too secure, and calls that the problem.

The story: Matthew Green published “Everything is about to go dark” on August 14. He wrote it after Usenix Security in his hometown of Baltimore. His prediction: within two years, major software runs out of remotely exploitable bugs. Entire CI toolchains are being rebuilt to scan with AI before a human touches the code. Defenders are grinding through decades of backlog.

Green thinks that’s bad, and the reasoning is second order. Law enforcement stopped fighting encryption in court once it could simply buy the hack. Take the exploits away and the demand for deliberate backdoors restarts, with nothing left to substitute for it.

The details:

  • He dates the turn to April 2026. Anthropic announced Mythos, a model unusually good at finding vulnerabilities, and the US government temporarily blocked its export. Green calls that restriction mostly pointless, since open weight labs including Z.ai showed nobody holds a monopoly on bug finding.
  • The last era has a start date. Apple encrypted iPhone storage behind the passcode in 2010. By 2016 nearly a billion WhatsApp users had default end to end encryption.
  • The paperwork fight is already live. In April 2025 the UK Investigatory Powers Tribunal refused to keep Apple’s technical capability notice case secret.
  • The EU runs a standing group on this. It endorsed 42 recommendations in May 2024, organized around data on the device, in the provider’s system, and in transit.
  • The only public comment disagrees. It argues most companies will add AI slop bugs faster than they remove real ones.

“The demand for constructed, intentional backdoors will re-start in earnest.” (Matthew Green, Johns Hopkins University)

Why builders care: If Green’s clock is close, AI scanning in CI stops being optional. Your dependencies get hardened while your own glue code stays exactly as soft as you left it. Expect the next access demand as a legal document, not a zero day.

🔐 PRIVACY, BILLED BY THE SECOND

A padlocked data cube feeding into a compiler funnel with a stopwatch hanging beside it

Google’s HEIR compiles a model you already trained to run on encrypted data.

The story: Google showed HEIR on August 14, an open source compiler for homomorphic encryption. The pitch on the project site is a build step. Write a program in Python, annotate which types are secret, and the compiler handles the rest. Doing that conversion by hand takes a team of cryptographers, which is Google’s stated reason for building it.

Four demo apps shipped with it, source in Google’s fully-homomorphic-encryption repo. A recommendation model, a credit card fraud detector, anomaly detection on encrypted network traffic, and a hotword detector.

The details:

  • The bill is in seconds, not milliseconds. The featured recommendation demo reports 24 seconds per inference on UCI and 228 to 489 on Criteo. Both on one CPU core.
  • The trend is the argument. Google’s 2023 announcement quoted 16 seconds for a 3-layer network. A full encrypted recommendation model now lands in the same neighbourhood.
  • One optimization made it affordable. The underlying HE-LRM paper reports a 56x speedup over prior art on encrypted embedding lookups, using client side digit decomposition.
  • Hardware is promised, not measured. Google names Belfort, Niobium, Cornami and Optalysys as accelerator partners, with no published latency numbers yet.
  • Nothing ships to users yet. HEIR is a compiler you run yourself. The post names no product currently running encrypted inference, and quotes no price.

”… it shifts the capability/privacy trade-off to a question of cost. And the cost of homomorphic encryption is rapidly decreasing.” (Jeremy Kun, Google)

Why builders care: Seconds per inference rules out chat. It rules in the jobs indie builders get locked out of today. Batch fraud scoring, overnight risk models, compliance checks for regulated buyers. Your sales line changes from a privacy policy to a guarantee the data never decrypts on your side.

📚 THE FAT ONE LOST

Two thick reference books on a desk, the fatter one pushed back into shadow

Mark Dominus bought the bigger thesaurus, then banished it to a farther shelf.

The story: Mark Dominus wrote up the seven books on the shelf he can reach without getting up. Not the ones he consults most. The ones he hopes will rub off on his writing. He bought Roget’s 8th edition and kept it beside his 1989 copy of the 4th. Same entries, looked up in both. Then he retired the 8th to a farther shelf.

His case for the book isn’t the word list. Roget sorted everything into a thousand divisions, so related ideas sit next to each other on the page. Leaf through the relative time run, from entry 116 Priority to 121 The Future. You refine what you meant instead of swapping in a fancier word.

The details:

  • The verdict on tools is the whole post. A thesaurus does nothing for a reader who doesn’t already know what the words mean. All that person can do is replace one wrong word with another at random.
  • Three public answers to the same problem. Dominus keeps 7 books within reach. Patrick Collison catalogued hundreds and hasn’t touched the page in around 10 years. Derek Sivers logs 482 books, each one rated.
  • The eighth slot is still empty. Retiring the fat one opened a space, and nothing in his office volunteered to fill it.
  • The base rate is worse than you’d guess. Pew found 25% of US adults read no book at all last year, against 14% who read more than 20.
  • The same shelf holds pure junk. Googly eye stickers, a box of 8mm backup tape from the 1990s, and a set of Korean playing cards.

“Effective tool use requires skill and training, and careful thought.” (Mark Dominus)

Why builders care: This is the failure mode of the AI assistant that hands you a bigger candidate list. The list only helps someone who can already judge the candidates, which is the part every tool pitch skips.

STACK OF THE DAY

🚨 Defencecore RLS audit

Defencecore published an agent prompt you hand to Claude Code or Codex. It walks every table in your Supabase project and checks the row level security policy. It went up as a Show HN with a single point. That’s the wrong score for the thing most likely to be quietly broken in your side project. Row level security is the Supabase footgun everyone knows about and nobody audits.

Not sponsored. We just feature tools builders would actually use.

BOOKMARKED TODAY

That’s the board for today. Go build something.

Curated by AI, built by a human.