The Diff

Archives
Log in
Subscribe
October 31, 2026

No. 028: A maliciously modified Claude Code build on Cursor's Marketplace planted an SSH backdoor. Nine AI platforms breached in October.

A maliciously modified Claude Code build on Cursor’s Marketplace planted an SSH backdoor, Wiz counts nine AI dev platforms breached in one month, and Binance’s $6.3B traces back to a single laptop. Plus the quiet ones nobody talks about.
View in browser: The Diff No. 028
The Diff  No. 028
Saturday, 31 October 2026  ·  AI dev news, only what is verified
Wiz Black October threat-update artwork
A maliciously modified Claude Code build on Cursor’s Marketplace planted an SSH backdoor, Wiz counts nine AI dev platforms breached in one month, and Binance’s $6.3B traces back to a single laptop.

Security week, day five, and the supply chain got personal. The extension marketplace your IDE trusts turned out to be an unsigned channel straight onto developer machines: for eleven hours on Thursday, “updating Claude Code” from the Cursor Marketplace meant handing an attacker a persistent key, a reverse tunnel and a list of your secret variable names. Both Cursor’s own notice and a full independent reverse-engineering are out, and the ten-second self-check is at the bottom of this issue. Run it.

Also today: Wiz connects nine October breaches at AI developer platforms into one “Black October,” Binance names its first root cause, Google Research puts LLMs to sleep so they stop forgetting, a sans-IO WireGuard lands in Rust, and somebody compiled an entire Nintendo into a stock Qwen.

One extension, one laptop, nine platforms. The self-check takes ten seconds.

IN THIS ISSUE
01
🔌 The official Claude Code extension on Cursor’s Marketplace shipped an SSH backdoor
For eleven hours on October 29, installing or updating “Anthropic.claude-code” from the Cursor VS Code Extension Marketplace planted a persistent SSH key, a reverse tunnel and a recon beacon. The payload fired on module load, before activate() ever ran.
02
🎰 Nine AI dev platforms breached in one month, and every attacker hunted the same thing
Wiz’s “Black October” report connects Fireworks AI, Andromeda, Zeabur, Lightning AI, OpenCode, Railway, Warp, CommandCode and Northflank: targeted exploitation going after stored environment variables, “with a focus on those granting AI access.”
03
💻 Binance now says the $6.3B started with one engineer’s laptop
The exchange’s first root-cause claim: a compromised DevOps laptop reached a service container and submitted the withdrawal transactions, with “no evidence of any wallet private key leak.” Withdrawals target a Sunday restart; the North Korea attribution is one firm’s claim.
04
😴 Google Research puts LLMs to sleep so they stop forgetting
The Sleep framework alternates “wake” phases that gather experience with offline “sleep” phases that consolidate it into the weights - claiming continual learning with long-context retention up to 10M tokens.
05
🦀 Two years of patching boringtun taught Firezone what WireGuard’s API should be
sanstun is a sans-IO, userspace WireGuard implementation in Rust: no internal buffering, rate limiting enforced by the type system, and packet encryption you can parallelize across cores.
06
🕹️ Someone compiled an entire Nintendo into the weights of a stock Qwen
nesformer runs the NES’s 6502 CPU, its 2KB of RAM and a cartridge ROM inside a standard Qwen3.5 checkpoint - no training involved, a compiler computed every weight. It plays Super Mario Bros. one predicted token at a time.
🔥 VIRAL
The loud ones on Hacker News. Numbers are a morning snapshot.
01  HN: 837 points at 06:35 IST · cursor.com + phorion.io, Oct 29-30 · no image available

🔌 The official Claude Code extension on Cursor’s Marketplace shipped an SSH backdoor

For eleven hours on October 29, installing or updating “Anthropic.claude-code” from the Cursor VS Code Extension Marketplace planted a persistent SSH key, a reverse tunnel and a recon beacon. The payload fired on module load, before activate() ever ran.
SCENE

Cursor’s own notice reads like an incident timeline, because it is one. Between about 2:27am and 1:35pm PT on October 29, two versions of the Claude Code extension - 2.1.312 and 2.1.313, Windows and macOS - sat in its marketplace with a payload inside. Cursor is explicit: “These malicious versions of the extension were not developed by Anthropic,” and it says there is no evidence other versions or the official distribution were affected. The marketplace has been shut down and rebuilt, keys rotated, and the attacker’s server taken offline with the cloud provider. Some coverage is already mis-framing this as “Anthropic hacked.” That is not what happened; this is marketplace poisoning under a borrowed name.

THE THING

Phorion’s reverse engineering shows what an update actually did. The payload runs when the module loads, before the extension’s activate() is called, so catching it afterward is not a defense. It appends an attacker Ed25519 key, comment svc-telemetry, to ~/.ssh/authorized_keys, opens a reverse SSH tunnel (-R *:0:localhost:22) to 134.209.82[.]225, persists through a com.apple.sysmond LaunchAgent or a SysMon scheduled task, and beacons home a host profile plus up to forty environment-variable names matching KEY, TOKEN, SECRET, PASS, AWS and GCP - names, not values, which reads like a shopping list for a second pass. The twist: the macOS enablement step is buggy (systemsetup -getremotelogin exits 0 on its failure path, so the check misfires), and on a Mac without Remote Login already on, inbound SSH never opens. But the keys and persistence land anyway, and Phorion’s warning stands: if Remote Login is on now or gets turned on later, “the backdoor is ready and waiting.”

✂️ STEAL IT

Steal the cleanup order, because uninstalling the extension is not enough: the persistence scripts keep running after removal. Remove the extension, delete the LaunchAgent or scheduled task, pull the svc-telemetry line out of authorized_keys, then rotate every credential that machine could see - from a different, trusted machine. Cursor published full IOCs: both VSIX SHA-256s, the key fingerprints and every file path.

Try today  VERDICT

Run the one-liner at the bottom of this issue on any machine that updated extensions this week; it takes seconds. It fails you as full reassurance: a clean grep does not clear the persistence locations, and a machine that had Remote Login on needs the whole IOC list, not just the key check.

Cursor’s security notice  ·  Phorion’s reverse engineering  ·  HN discussion

Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes.
More like thisLess like thisDig deeperAlready knew it
Wiz Black October threat-update artwork
02  Northflank on HN: 390 points at 06:35 IST · wiz.io, Oct 30 · image: Wiz

🎰 Nine AI dev platforms breached in one month, and every attacker hunted the same thing

Wiz’s “Black October” report connects Fireworks AI, Andromeda, Zeabur, Lightning AI, OpenCode, Railway, Warp, CommandCode and Northflank: targeted exploitation going after stored environment variables, “with a focus on those granting AI access.”
SCENE

The ninth incident landed yesterday morning. At 11:43 UTC on October 30, a Wiz threat alert fired at Northflank, a platform that builds and hosts your services - and stores the environment variables that come with them. By evening, its status page told every customer to “assume their environment variables are decrypted and begin rotation immediately,” and reported that “multiple users are confirming that their AI keys have been used.” Overnight, audit logs were switched on for everyone, free tier included. Sentry published its own customer-impact note, because a PaaS breach is never one company’s breach.

THE THING

Wiz’s report is the first to say the quiet part: these are not nine unlucky companies. Fireworks, OpenCode and Warp all report indications of an “agent swarm” attack; OpenCode and CommandCode attribute the motive to attempted model “distillation”; Zeabur is shutting down entirely over its incident. The case study is LLMJacking at scale: a stolen key validated with a single InvokeModel call within thirty minutes of the breach - it asked for Claude Fable 5.1, failed, and settled for Opus 5.5 - followed by roughly 800,000 Bedrock calls over 30 hours, 90% of them successful, routed through global.* cross-region inference profiles so only about 6% ever throttled. Three quarters of the spend was Opus. Somebody’s stolen key had a very productive day.

✂️ STEAL IT

Steal the detection shape: a first-ever InvokeModel from a new principal, followed immediately by sustained cross-region inference traffic, is the attack working as intended, not a misconfiguration. The case-study dwell time - 800k calls before anyone stopped it - is the number to beat with your own alerting.

Watch  VERDICT

Read the report if your platform stores other people’s secrets; October’s nine names are probably not the whole list. It fails you as attribution: Wiz is a vendor describing its own telemetry, no shared CVE links the nine incidents, and nobody can name the actor yet.

The Wiz report  ·  Northflank’s status page  ·  HN discussion

Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes.
More like thisLess like thisDig deeperAlready knew it
03  Wired on HN: 376 points at 06:35 IST · binance.com + CoinDesk + NPR, Oct 30 · no image available

💻 Binance now says the $6.3B started with one engineer’s laptop

The exchange’s first root-cause claim: a compromised DevOps laptop reached a service container and submitted the withdrawal transactions, with “no evidence of any wallet private key leak.” Withdrawals target a Sunday restart; the North Korea attribution is one firm’s claim.
SCENE

The day after the drain, the story narrowed from fifty blind signatures to one machine. Binance’s updated announcement says the incident “may involve a compromised DevOps engineer’s laptop, which allowed an attacker to gain access to a service container and submit unauthorized withdrawal transactions,” and adds that it found “no evidence of any wallet private key leak.” The tally now stands at around $6.3B impacted, about $3 billion of it frozen on-chain, with SAFU and company funds covering all losses. Co-CEO Richard Teng says investigators found additional breach points beyond the laptop, and that the attackers “used a self-destruct mechanism to erase evidence.”

THE THING

Two more pieces landed alongside. The attribution: a U.S.-based blockchain intelligence firm told NPR that a North Korean unit was likely behind the hack - one firm’s assessment, echoed by CoinTelegraph naming TraderTraitor, not an established fact. And the timeline: Wired’s reconstruction has the whole operation “pulled off in just 20 minutes,” which does not contradict the 204-second drain window No. 027 reported - that was the gap between the fifty authorizations and the sweep, while Wired counts the full operation. Withdrawals stay suspended; the restart target is Sunday, and Teng says service “won’t be switched back on until it passes rigorous testing,” with an updated Proof of Reserves promised by Monday.

✂️ STEAL IT

Steal the honest sentence for your own incident reports: “may involve a compromised DevOps engineer’s laptop” names a mechanism and keeps the uncertainty attached. The gap that matters is still open: how container access turned into fifty hot-wallet signatures is exactly what the promised post-mortem has to explain.

Watch  VERDICT

Watch for the full post-mortem and Sunday’s restart; both were promised, and neither had landed at publication time. It fails you as closure: the root cause is the exchange’s initial claim, and the attribution rests on a single firm’s assessment.

Binance’s announcement  ·  NPR on the attribution  ·  Wired’s reconstruction  ·  HN discussion

Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes.
More like thisLess like thisDig deeperAlready knew it
🔭 UNDER THE RADAR
Quiet tools and field reports worth your time.
Sleep framework cover art, Google Research
04  HN: 186 points at 06:35 IST · research.google, Oct 27 · image: Google Research

😴 Google Research puts LLMs to sleep so they stop forgetting

The Sleep framework alternates “wake” phases that gather experience with offline “sleep” phases that consolidate it into the weights - claiming continual learning with long-context retention up to 10M tokens.
SCENE

Catastrophic forgetting is the tax every continually-trained model pays: teach it something new, watch an old skill erode. Google Research’s Ali Behrouz and Vahab Mirrokni propose fixing it the way brains apparently do. Their Sleep framework alternates active interaction with offline memory consolidation, and they claim it “allows LLMs to continuously acquire new facts, retain long-context knowledge up to 10M tokens, and improve complex reasoning.”

THE THING

The mechanism borrows more than the metaphor. During wake phases the model keeps a fast, short-term memory of new experience. During sleep, it unlocks pre-allocated Mixture-of-Experts memory blocks and consolidates through Generalized Knowledge Distillation plus RL imitation with a Levenshtein reward; then a “synaptic pruning” step resets the temporary parameters so the next wake phase starts clean. “Dreaming” is literal in the implementation: random expert activation generates synthetic rehearsal data, filtered by gradient signal, so the model practices on self-generated variations. The reported wins - against in-context learning and SFT baselines, on class-incremental learning, continual translation and a 10M-token long context - are all Google Research’s own.

✂️ STEAL IT

Steal the pattern even if the numbers only hold in-house: a fast scratchpad, a slow consolidated store, and an explicit reset between them is an architecture any agent builder can copy this week. Your agent’s “sleep” is a cron job that compresses yesterday’s logs into memory.

Watch  VERDICT

Watch for independent replication before designing around the 10M-token claim. It fails you as settled science: one team, its own baselines, and a blog-post venue rather than peer review.

The post  ·  HN discussion

Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes.
More like thisLess like thisDig deeperAlready knew it
05  HN: 91 points at 06:35 IST · firezone.dev, Oct 31 · no image available

🦀 Two years of patching boringtun taught Firezone what WireGuard’s API should be

sanstun is a sans-IO, userspace WireGuard implementation in Rust: no internal buffering, rate limiting enforced by the type system, and packet encryption you can parallelize across cores.
SCENE

After more than 70 patches to their boringtun fork, Firezone’s Thomas Eizinger writes, they finally knew what the API should have been all along - so they built it. sanstun, announced yesterday, is a from-scratch, sans-IO implementation of the WireGuard state machine in Rust. Sans-IO means the library owns zero side effects: no sockets, no threads, not even time or randomness. That is what makes it easy to test, easy to embed, and free of function-coloring fights with the rest of your codebase.

THE THING

Two design choices carry the post. First, every inbound packet must enter through Interface::receive - the only public parser - which applies cookie rate limiting before any expensive Diffie-Hellman work, and the handshake-type constructors are crate-private, so forgetting the rate limiter is unrepresentable rather than merely discouraged. Second, encryption is split from bookkeeping: Tunnel::begin_encapsulate reserves a packet counter and returns a PendingEncapsulate you can hand to another thread, which is how a single tunnel’s flows parallelize across cores without tripping WireGuard’s replay counter.

✂️ STEAL IT

Steal the crate-private-constructor trick for any security-critical path: if the dangerous call can only be reached through the safe wrapper, the invariant stops depending on the reviewer’s memory.

Watch  VERDICT

If you maintain VPN or tunnel code in Rust, the design tour alone is worth the read; adoption is a bet on a day-old library. It fails you as proven software: announced October 31, no independent audit, and boringtun still holds all the production miles.

The announcement  ·  HN discussion

Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes.
More like thisLess like thisDig deeperAlready knew it
Super Mario Bros. World 1-1, played by the weights
06  Show HN: 213 points at 06:35 IST · github.com/dbreunig · image: the project

🕹️ Someone compiled an entire Nintendo into the weights of a stock Qwen

nesformer runs the NES’s 6502 CPU, its 2KB of RAM and a cartridge ROM inside a standard Qwen3.5 checkpoint - no training involved, a compiler computed every weight. It plays Super Mario Bros. one predicted token at a time.
SCENE

Two weeks after No. 026 covered Qwenputer - a RISC-V CPU compiled into Qwen’s weights - Drew Breunig went and put a whole games console in there. nesformer builds directly on Qwenputer’s compiler and its central trick, that the model’s Gated DeltaNet state is the machine’s memory. The NES’s 6502 CPU, its 2KB of work RAM and a cartridge’s program ROM all execute inside a stock Qwen3.5 model; a small host program draws the picture, reads the controller and fires the vblank interrupt around it.

THE THING

The demo is Super Mario Bros. World 1-1, played by the weights. The model predicts the CPU’s execution log one token at a time, and for the bundled test games every frame matches a conventional 6502 emulator’s output exactly - the author’s own check, with the method documented in the repo. Nobody trained anything; the compiler computed the weights the way it would compute machine code. Weights are quietly becoming a compilation target.

✂️ STEAL IT

Steal the demo idea for your next model-format pitch: “it runs Mario” communicates “this representation is exact and general” faster than any architecture diagram ever will.

Try today  VERDICT

The repo bundles the weights and the games; if you have the disk space, this is the most fun you will have verifying someone else’s checksum. It fails you as more than a marvel: it is a spectacularly slow way to emulate a 1983 console, and frame-exactness is the author’s own measurement.

The repo  ·  HN discussion

Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes.
More like thisLess like thisDig deeperAlready knew itTest it for me
📊 THE ATTENTION GAP
HN points at 06:35 IST. Attention, not quality.
Claude Code extension backdoor
837 HN
Northflank breach
390 HN
Binance in 20 minutes (Wired)
376 HN
Proton re-interviews
266 HN
A Nintendo in Qwen weights
213 HN
ALSO ON THE RADAR

Proton’s CEO cancels the AI “re-interviews.” After 404 Media’s investigation into Andy Yen’s aggressive AI push - including his “that’s bullshit and loser thinking” reply to an engineer who asked for a few days to pentest AI-written code - Proton cancelled the near-company-wide re-interviews that would have graded staff on AI use. Yen told the all-hands “I’m still going to push,” employees say the direction is unchanged, and his Reddit response has passed 1,200 downvotes. The report →

Cisco hands a 27-year-old unlimited tokens and 98% of its legacy code. President Jeetu Patel found ex-founder Padmanabhan Krishnamurthy “buried” five levels down and gave him four weeks to prove a two-person, AI-augmented team can refactor Cisco’s legacy estate - goal: 98% within 24 months. “Go nuts,” per Fortune’s profile. The outcome is very much not in yet. The profile →

BeOS now runs in your browser. Infinite Mac added the road-not-taken OS Apple once considered buying: full BeOS installs that boot instantly from streamed disks, mount Internet Archive CD-ROMs and keep a Saved HD. Most of the work went upstream to DingusPPC - and the author notes an AI agent helped with the fiddly reverse engineering. The post →

cfsolve cracks Cloudflare Turnstile without a browser. A Rust library and CLI that returns a Turnstile token or a cf_clearance cookie for widgets and challenge pages, no headless browser required. Useful for testing your own site; obviously useful to everyone else too. The repo →

Follow-up: Alpine’s GitLab is back. Rebuilt from scratch on GitLab 19.3.7 with the CVE-2026-97229 fix, all passwords and 2FA reset, new SSH host keys. Alpine says it is already seeing automated logins with passwords leaked from other sites - if you reused yours there, change it everywhere else it works. The post →

Follow-up: PAN-OS hotfixes are all out. Palo Alto’s advisory now lists fixed builds for every affected train (12.2.4-h1, plus the 12.1.x, 11.2.x and 11.1.x hotfixes) for the PHP2Shell pair from No. 027. If your management interface answers the internet, that issue’s audit still applies. The advisory →

Follow-up: eMudhra’s macOS gap is still open. Nothing new on the disclosure page since October 28; Apple’s bundled git, curl and python3 still trusted the compromised intermediates as of that update. No. 026 has the story and the one-minute self-check. The status page →

🗂️ THE BORING STUFF

Oct 30 · GitHub’s Actions, Pages and Webhooks incident is resolved. A Thursday-evening UTC incident hit Actions runs, Pages deploys and webhook deliveries; GitHub’s status page marks it resolved. Relevant if your evening deploys looked haunted - they were. GitHub status →

Oct 29 · Polymarket halted international trading for an hour after a vendor compromise. NBC reports a third-party provider (identified by a source as analytics tool DebugBear) was compromised; some users reported drained wallets, The Block cites roughly $2.65M via Arkham, a figure Polymarket declines to confirm. The company says everyone affected will be “made whole.” The U.S. platform was not affected. NBC →

Oct 29 · OpenAI doubles ChatGPT Pro usage limits after backlash over a planned 50% cut. Announced on X and reported by BleepingComputer: Pro 100/200/500 tiers and Business Premium get doubled limits, the planned Pro 200 reduction is off, and Dots usage stays separate. Exact numbers were not disclosed. BleepingComputer →

Official pages and named outlets read this morning. Nothing installed or run; their numbers are theirs.
🧪 ONE THING TO TRY TONIGHT
Check your own keys for svc-telemetry.
Free and ten seconds: run “grep -qs svc-telemetry ~/.ssh/authorized_keys && echo FOUND || echo clean” on every Mac and Linux box that could have updated the Claude Code extension on Thursday - CI images and shared dev machines included. A clean result is necessary, not sufficient: Cursor’s IOC list has the LaunchAgent, the scheduled task and the file paths to check next, and any hit means rotating that machine’s credentials from a different computer.
What this issue could not see. HN, Lobsters and GitHub were swept overnight and this morning, and primary sources were read: cursor.com, phorion.io, wiz.io, status.northflank.com, binance.com, coindesk.com, npr.org, wired.com, research.google, firezone.dev, github.com/dbreunig/nesformer, github.com/chris124567/cfsolve, 404media.co, fortune.com, blog.persistent.info, bleepingcomputer.com, nbcnews.com, alpinelinux.org, security.paloaltonetworks.com and githubstatus.com. Reddit is absent again: anonymous access from hosted networks is blocked by policy and the official app-registration route is unresolved, so no Reddit source was consulted. X blocks anonymous access, so quotes from X posts are as relayed by the cited outlets. The North Korea attribution is one blockchain-intelligence firm’s assessment reported by NPR; Binance’s laptop root cause is the exchange’s initial claim, with the full post-mortem still pending at publication; the Wiz report is a vendor describing its own telemetry, and no shared CVE or named actor links October’s nine incidents; Cursor’s containment account is Cursor’s; the Sleep framework’s numbers are Google Research’s own, from a blog venue; nesformer’s frame-exactness is the author’s check; Fortune’s Cisco story is a profile of an internal program, told by its sponsor. Nothing was installed or run. Deduped against every published issue, No. 001-027.
Talk to the editor 💬

Reply to this email with anything: what to chase, what to drop, what you want more of. Every reply shapes the next issue.

More of this kind of issueLess of this
Images: credited in each report; source screenshots and announcement artwork, not our tests.
Don't miss what's next. Subscribe to The Diff:
← Newer No. 029: One shared link took over ChatGPT accounts. Malware hired an LLM analyst. Older → No. 027: Fifty signatures, 204 seconds, billions gone. Mistral opened its flagship.
Powered by Buttondown, the easiest way to start and grow your newsletter.