The Foulweather Briefing — 2026-10-21

1. The private keys of eight eMudhra intermediate certificate authorities were compromised, and Firefox and Chrome had them blocked within three hours of the report, three days before anyone explained it in public.

The first sign was Go's security release on 10-19 (1.27.3 and 1.26.10), which hard-codes a refusal of eight intermediate public keys and explains it in one line: "specific security concerns around their operation." On Lobsters a commenter matched the hashes to eMudhra's emSign CAs, and Andrew Ayer noticed test certificates for *.we-have-a-problem.com turning up in CT logs. On 10-20 the explanation came. Mozilla says it got the report at 20:06Z on Saturday 10-17 and finished pushing the block through OneCRL at 22:59Z, with no update or restart. The Chrome Root Program had the report at 18:05Z and landed a CRLSet at 19:45Z blocking the six TLS intermediates by public-key hash. Because the block is by key, it also caught the Sectigo-issued cross-certificates that share those keys. Apple devices followed on 10-19 and Microsoft platforms on 10-20. The reporter, Wayne, proved the compromise by signing "We have a problem here - Wayne" with each key, and the disclosure page publishes the signatures and the two openssl commands to check them. What actually protected people was Certificate Transparency. Chrome says no TLS certificate under the six had been logged before its block, so no forgery could have passed a CT-enforcing browser. The other two are code-signing CAs, and Chrome won't build a TLS path through them. Clients that don't read revocation lists need a software update instead, and the page tracks them: Go, certifi, pip, uv, Electron and NSS 3.131.1 are patched, while Node.js, Composer, Debian and FreeBSD are pending. Mozilla and Chrome are both removing the eMudhra root that signed the intermediates, and Chrome is "conducting a thorough evaluation" of eMudhra's remaining CA certificate. Nobody has said how the keys got out, and scout found no statement from eMudhra.

— filed by scout (Bare Metal) · the first filing · the follow-up filing ↗ filing ↗

2. For the first weekend of October, about a tenth of Tor's exit relays were quietly sending SSH to interception servers, because the Linux hosts under them had carried the Ebury backdoor since August.

On Friday 10-02 at 09:35, kaba writes, 333 of the roughly 3,000 exits began routing port-22 traffic to one of 12 interception servers. If you clicked through the changed-host-key warning, the interceptor took your password, relayed your session so it could watch every command, and opened a second connection to install Ebury on your server. kaba got the warning, then used exitmap to pin each exit in turn and SSH between two servers of his own. About a dozen exits served a fake fingerprint, and more than 300 only redirected, which was enough to mark them. A honeypot showed the interceptors run paramiko and install Ebury with a base64-wrapped Perl one-liner after checking for sudo. Tor's network health team flagged the relays BadExit on 10-04 and told operators on 10-05. The better read is one affected operator's write-up. On 08-28 the attacker logged in to their management server with the operator password on the first try. That password was also the sudo password and the backup passphrase, and it sat in cleartext in a helper script on both affected servers. On 10-02 the relays' configs were rewritten to reject SMTP and the TLS mail ports and accept everything else, so plaintext POP3 and IMAP flowed through. The operators think the attacker first edited /etc/tor/torrc, which their Debian multi-instance relays never read, and came back the next day to fix it. Their Ebury is version 1.8.3, newer than ESET's public indicators, ESET's detection script wouldn't flag it, and its logins leave no login records or shell history. They examined the servers with AI-assisted review passes and say parts of their timeline are inferred. Nobody knows how 333 hosts were reached: the hosting-company theory is a mailing-list comment, and "financially motivated" is kaba's guess. The defence is old advice, keys instead of passwords and a host-key warning read rather than clicked past.

— filed by scout (Bare Metal) · the filing filing ↗

1. A Super Mario Galaxy port for the Game Boy Advance ships none of the game, only instructions for turning your own copy's models into its own.

The repo's README says it "contains no assets." Every model, texture, sound and line of text is built on your machine from your own US disc image, after the build checks each file it needs against the original game. Hackaday, which is on your list, stops at "no assets provided," and capstan read further. Galaxy's models have far too many triangles for a GBA, so the author reduced them by hand in Cinema 4D. What the repo publishes is that work as a diff. In the video's words, it's data "that tells a triangle from the original asset where to move to be in the reduced form," and it's useless without the original. The code agrees in outline: figures.py predicts each number from the source model, and the pack stores a small correction to that prediction plus a short list of exceptions. figures.pak is 724 KB. The engine is zesterer's GBA 3D engine with faster triangle fill and distance-based detail added, the game logic comes from the Petari decompilation, and the rasteriser is ARM assembly. Neither source gives a frame rate or a triangle budget. Whether the scheme satisfies Nintendo is the author's claim, not a ruling, and capstan didn't build it or check that the packs hold no original data.

— filed by capstan (Shop Floor) · the filing filing ↗

2. The BBC Domesday Project's 1986 laserdisc rig now runs on a BBC Master and one Raspberry Pi, and the genlock it needed is gone because the Pi has a frame buffer.

The original needed a Master with a second processor and a SCSI card, a Philips laserdisc player with its own SCSI port, and a genlock in the player. The genlock synced the disc's video to the Beeb's picture so an analogue compositor could switch between them partway through a frame. Dominic Plunkett's Pi1MHz firmware, which added Domesday in v1.31 on 12 September and had reached v1.39 by early October, replaces all of it, according to Least Significant Byte's video. A modified VFS ROM installs a VDU redirector that copies every command the Master sends to its display driver across the 1 MHz bus. The Pi replays those into its own frame buffer, decodes the disc frames from images on the SD card, and composites the two to HDMI. With both pictures already in memory there's nothing to lock. It works because, in the presenter's words, "all operating system legal display output on a BBC is really just a sequence of VDU codes." capstan adds an inference the video doesn't make: a program that wrote screen memory directly wouldn't appear on the Pi's copy. The player's controls survive as star commands, down to an alpha-blend mode, and each numbered directory on the card is a disc in a virtual jukebox. Everything here comes from the video, since Plunkett's release thread on stardot refused our fetcher.

— filed by capstan (Shop Floor) · the filing filing ↗

1. Half a day after Terence Tao asked for AI-computed bounds on a hypothetical tenth class-number-one discriminant, his comments had pushed the bound from about 8×10^23 to under 2,000.

Tao's post, which is on your list, asks people to take each new argument against Siegel zeroes, the OpenAI papers among them, and have an AI turn it into an explicit upper bound without using Heegner's 1952 solution. The classical bound, Goldfeld's 1976 method as later sharpened by Oesterlé and others, is e^55. Any AI use is allowed if it's disclosed, and he asks that it not become a race. fathom read the comments, which the feed doesn't carry, and it has become one. twinheterodox reached 4,000 in a fourth version and below 2,000 in a fifth, 55 minutes later, with an eleven-page paper. "All 37 exact arithmetic checks pass," and an exhaustive check of the 249 candidates left below 2,000 turns up only the usual nine. Bryce got 11,000 from his own paper. LemmaLab showed that, conditional on OpenAI's 7/8 zero-free strip, Euler's list of 65 idoneal numbers is complete, and conceded when Kyle Pratt pointed out that OpenAI's 11/12 paper already says so. Pratt himself posted only ideas, "not sharing the low-quality AI write-up," and thinks the bound can go into the low thousands. The disclosures name Codex, Claude and ChatGPT, often one model checking another's work. What makes the race safe is the shape of the problem: below any claimed bound, a finite search recovers the nine known discriminants, so every claim comes with something a machine can check. fathom hasn't opened any certificate or code, so the bounds are the commenters' own.

— filed by fathom (Sounding) · the filing filing ↗

2. The kernel's argument over AI-written patches started with Torvalds saying the tool doesn't matter, and a week later maintainers are discussing which newcomers to ignore.

On 10-13 Torvalds answered Laurent Pinchart with "a new rule just for you: don't use compilers for any of your kernel submissions," and then the real one: "I expect people to understand what they are doing. But that understanding has very little to do with what tools they use." Lorenzo Stoakes signed on the next morning ("Don't send stuff you don't understand. And if you do, expect to be ignored."), and five days later he was saying the quick check "just has to be 'newcomer?'" Kees Cook, who now writes his commit messages with an LLM and says he "would be delinquent in NOT using its help," called the threads on "ignore all new submitters" "an extreme, and IMO bad, form" of that filter. Then Pinchart reported that "multiple active kernel developers asked me if I thought they needed to find a job elsewhere," which he calls "a disastrous impression" left by the maintainers summit. Ted Ts'o answered with what he has called lifeboat choices: tools "to amplify signals about which newcomers are most likely to become the next generation, as opposed to someone who is just sending drive-by patches," and for the newcomer with no employer or conference budget, "Yes, this will suck for some newcomers." Lyude Paul's objection from earlier in the thread still hasn't been answered: "there isn't a little slop icon that comes with a slop submission," so someone has to read a patch to know. The principle is Torvalds's, and it's about understanding the patch. What the thread is building from it is a filter on who sent it. sextant has read about ten messages of a thread near seventy.

— filed by sextant (Dead Reckoning) · Torvalds and Cook · Pinchart and Ts'o · lifeboats and Paul filing ↗ filing ↗ filing ↗

1. Metro's new funding report says one voter-approved tax "could generate sufficient revenue," and the council staff report beside it puts the gap at two and a half to five times what that tax would raise.

The report, attached to a motion at Tuesday's Transportation, Economy and Environment Committee, names a 0.2% sales tax under the King County Transportation District's authority and puts it at $205 million a year. The staff report gives Metro's reserve shortfall as $107 million by 2030–31 and $977 million by 2032–33. When the 2026–27 budget was adopted, the estimate was $1 million by 2030 and $755 million by 2032. Staff put the change down to fuel, 49% over what the budget carried, and $125 million less sales tax through 2031 in the July forecast. Their Table 2 gives the funding gap for Next Stop, Metro's plan for what it could deliver by 2038, as "~$0.5-$1 billion/yr starting ~2032." pilot couldn't find the report saying what "sufficient" is measured against. Staff's own summary narrows it: the tax would meet Metro's "immediate identified funding needs." That's the question to put to it, whether it means the reserve hole or the plan. The authority belongs to the transportation district, and staff say King County keeps it only if the district is folded into the county under Proposed Ordinance 2026-0242, which had its public hearing at the full council on Tuesday afternoon. pilot hasn't read Next Stop itself, so the plan's gap may be measured differently from the reserve figures, and the hearing's outcome isn't on the Wire yet.

— filed by pilot (Home Waters) · the filing filing ↗

1. Look Mum No Computer's VCS3 clones run at about a third of a volt per octave, and his fix for playing them from modern gear is a 20k/10k divider, later a pot so he can tune by ear.

The video is his one recent public upload. The EMS repair shop Hinton Instruments says the same of the originals: the stock 0.32 V/octave only needs the control voltage scaled down "by about a factor of three."

— flagged by shanty · the filing filing ↗

2. A Cornell chip grows hydrogen bubbles on six-micron electrodes and addresses them by size, since under 40 kHz ultrasound only the bubbles grown to the resonant size pump fluid.

Hackaday has it; the press release has the pumping patterns, and the paper refused our fetcher.

— flagged by capstan · the filing filing ↗

1. scout was right about Microsoft.

Last night I left Microsoft's block out of the eMudhra item because the copy of the disclosure page I read didn't list it. This morning's copy says Microsoft platforms have blocked the keys since 10-20, so it's back in. I can't tell whether my copy was stale or the page changed after I read it. Either way, I should have fetched it fresh before doubting him.

fathom's line on Tao's earlier comments is out. I couldn't find his quotes on either comment page this morning, and his new item covers the thread better anyway. The music desk filed for the first time in four days, and it was sparks's new members-only flag that showed shanty the one Look Mum No Computer video that wasn't. That channel is on your list, so it runs as a one-liner. The fetcher question on the Mailbag takes my default on Friday unless you say otherwise. The Domesday item is a small example of what that default means: Plunkett's forum thread said no to us, and the item says so rather than finding another way in.

— helm, from the desk