an ai found 3 bitcoin core bugs. bitcoin now goes public in days, not a year

daily ai × crypto news in 5 minutes. no hype, just what matters.

evening. an ai scanner went through bitcoin core, the software most bitcoin nodes run, and found three bugs. the worst could crash any windows or mac node from the outside. it went from "ai found it" to "public" in five days, and that speed is the real story: bitcoin core now treats bugs an ai could find as already known. also: bitcoin miners have signed $100b+ of ai deals but bill only about $1.1b a year, a new standard lets ai agents pay for things by mining, and agents just got trading keys that can't withdraw (but can still lose).

what the ai found. on oct 28, bitcoin core published three security advisories at once: two medium-severity bugs and one low. all three were first found by an automated llm scan run by project loupe, elias rohrer's project at spiral, and validated by bitcoin core developer gregory sanders. fixes shipped in four releases for every maintained version: 29.5, 30.4, 31.2 and 32.1. (bitcoin-dev mailing list, advisory)

the worst one. on windows, macos and bsd, bitcoin core watches its network connections with an older method than on linux. a malicious peer could send "tcp urgent data", a rarely used kind of network packet, and slip past the limit on how much data a node queues per peer. in the advisory's words, a malicious peer could "trivially crash any windows, macos or bsd node it is connected to". linux nodes were never affected. that's a denial-of-service bug: it knocks a node offline, it doesn't touch coins. timeline: loupe flagged it on oct 23, sanders validated it and told the security team the same day, the fix merged oct 26, releases came oct 27–28, and public disclosure on oct 28. five days, ai to public. (bitcoin core advisory)

the human was six days behind. the second bug: a peer could pretend to run an ancient protocol version, stop reading, and make your node queue messages for it forever, growing memory by up to 2 mb an hour per inbound peer and 5 mb an hour per outbound one. loupe reported it on sep 15. on sep 21, developer l0rinc independently found the same flaw and was told it was already reported. the third bug, a low-severity "undo file" overflow that's impractical to exploit (an attacker would need roughly a month of the entire network's block production), was reported three times: by loupe on jul 8, by l0rinc on aug 21 and by andrew dorman on oct 23. one catch: nodes on 32-bit systems or with blocks stored on fat32 drives stay exposed to that one even after upgrading. (advisory 2, advisory 3)

the rule that changed. normally, bitcoin core keeps medium and high bugs secret until two weeks after the last affected version goes end-of-life, which the project itself puts at "approximately a year" after the fix ships. that gives nodes time to upgrade before attackers learn what to look for. but its policy now has a separate class: "easily discoverable" bugs, defined as ones "that can be discovered with minimal effort, such as through a simple llm prompt". for those, it ships fixed releases "as soon as possible" with an advisory, skipping the embargo. all three of these went through that fast lane. three other bugs pre-announced with the latest release follow the normal process, with details due in about two weeks. (security policy, mailing list)

why it matters. the old logic of quiet patching assumed attackers were slow to find bugs. if a model can find a bug in an afternoon, the attacker's model can too, and secrecy only protects the attacker. so the honest move is: patch, publish, tell people to upgrade. we saw the attacker's side of this in php2shell, ai-found flaws that put a bitcoin dare server at risk in #013. this is the defender's side, and every cve that comes out of an ai scan will put the same pressure on other crypto projects. none of the three could steal coins or split the chain. if you run a node on windows or macos, upgrade. if you run it on linux, upgrade anyway. more on the network: bitcoin

bitcoin miners signed $100b+ of ai deals. about $1.1b a year is actually billing. coinshares' q2 mining report says more than $100b of disclosed ai/hpc backlog supports only around $1.1b of annualized revenue, with roughly 550 mw billing against more than 4 gw contracted. the market pays anyway: miners with contracted ai capacity trade at 12.9x ev/sales, versus 3.7x for those without. (coinshares q2 report)

why they're switching: at the end of q2 the listed miners' average cash cost to produce one bitcoin was about $75,500, while btc ended the quarter at $58,400. ai work earns about $1.5m per megawatt a year versus roughly $500,000 for mining, coinshares' bill cannon said on oct 29, and miners reported $512m of ai/hpc revenue in q2, about 32% of their total. iren's ai cloud revenue ($70.5m) overtook its mining revenue ($66.7m) for the first time, and it expects its exit from mining to be substantially complete by dec 31. (coinshares q2 report, etf database)

and they may not come back: fitting a site for ai tenants costs about $10–15m per megawatt versus $1–1.5m for mining, per coinshares' alex schmidt. the bet is on converting contracts into bills. for how this started, see our hashrate and miners' ai pivot issue; miners renting out compute are turning into neoclouds. (etf database)

a new way for ai agents to pay: by mining. quai network launched "x402 work" on oct 30. it builds on x402: when an agent or browser asks for something paid, the server answers "402 payment required" with a price and a small proof-of-work puzzle. the device solves it with its own cpu, gpu or asic and gets access, with no account, wallet or api key. the first service on it, hashference, offers 45+ ai models through openai- and anthropic-compatible apis, payable by mining, crypto credits or per-request proof-of-work. (altcoin buzz, x402 work, hashference docs)

the fine print: "pay with electricity" is still paying, with your own power bill. solving the puzzle doesn't earn a block reward unless it also meets the network's mining target, and the use cases listed (agents buying data, publishers charging scrapers) are, in altcoin buzz's words, "intended use cases rather than evidence of widespread adoption". our x402 explainer covers the standard it builds on. (altcoin buzz)

agents got trading keys that can't withdraw. they can still lose everything. tria, a self-custodial app that says it serves 600,000+ users, launched agentic trading on oct 30: connect claude, chatgpt, codex or cursor and let it trade perps on hyperliquid and decibel from your own account. the agent gets a dedicated trading key that can open and close positions but can't withdraw, and you can set it to expire or revoke it. tria's own risk note is the honest part: trade-only permissions "do not prevent trading losses". an agent that can't steal your money can still liquidate it, and in #015 we saw 9 of 10 models raise their own spending caps. (chainwire press release)

same day, in abu dhabi, 200+ people at xdc network's "unsupervised" build day let ai agents spend usdc on real coffee and merchandise; agent payments passed $1,000 by the end of the day. (cryptocompass)

project loupe: the ai scanner behind today's big one. an open-source (apache-2.0) harness from spiral and block that runs llm agents over a codebase, makes each agent prove its finding with a regression-test proof-of-concept, and reports only confirmed bugs (github). pricing: free; you bring your own model access and tokens, and spiral and block fund the scans they run for open-source bitcoin projects (launch post). for maintainers and security teams of open-source bitcoin software. more picks in the gweiwire tools catalog.

the other three bugs. bitcoin core says three more vulnerabilities, pre-announced with its latest release and unrelated to today's, get full details in about two weeks through the normal process. watch whether those were found by humans or by ai, and whether other crypto projects copy the "easily discoverable" fast lane.

that's the wire. if your node runs on a mac, update it tonight. back tomorrow. — gweiwire

forward this to a friend who runs a node in the corner of their room. new here? subscribe at gweiwire.com.