· 15 min read
Binance's $6.43B theft began with a fake AI installer; Postman rewired Codex and Claude Code
Published
Binance's post-mortem traces the $6.43B theft to a trojanized AI-tool installer, a fake job interview and a hijacked internal AI tool. Developers caught Postman installing itself into Codex and Claude Code, and Google says stolen LLM keys have run up bills of up to $1M.
Today in 30 seconds
- Binance says the $6.43B theft was a months-long campaign: a trojanized AI-tool installer, a fake job interview and a hijacked internal AI tool.
- A GitHub issue shows Postman desktop 12.33.4 silently installed plugins into Codex and Claude Code and approved its own startup hook.
- Google says stolen LLM keys are resold through "LLM mixers", with bills of over $100,000 a day and up to $1M per victim.
- Anthropic's open-source FlashProve lets anyone check cryptographically which model produced an answer, at roughly 70–90x more compute.
- A Stanford study finds 44% of 50 widely cited LLM findings no longer hold on newer models.
Free. Daily, 10-min read. Confirm by email, unsubscribe in one click. Privacy
Binance published how attackers got from a fake “AI tool” installer and a fake job interview to $6.43B of its crypto, with a hijacked internal AI tool in the middle of the chain. Developers also caught Postman quietly installing itself into Codex and Claude Code.
Google says stolen LLM keys are running up bills of up to $1M. On the lighter side, yesterday’s weight-editing trick now works on a laptop model, and Anthropic showed a way to prove which model wrote an answer.
Top news
Binance’s post-mortem: a trojanized “AI tool” installer, a fake job interview and a hijacked internal AI tool led to the $6.43B theft
Binance says the record theft was a months-long, multi-stage campaign, not one phishing click.
It started in June with an engineer downloading a trojanized Homebrew installer to get a popular AI tool. It ended on 29 October with a malicious container talking to an internal wallet-signing service.
- June 2026: a Wallet DevOps engineer (“Engineer A”) searched for a popular AI tool, landed on a trojanized Homebrew installer and ran it.
- Binance caught it and rotated credentials, but details about his role and his wallet-team colleagues had already leaked.
- September: a second engineer (“Engineer B”) was approached on a real job-hunting site by a fake “Project Director”.
- After an HR call and a “CEO” interview on a chat app, he was told to download an “assignment” repository onto his work computer.
- Its setup step ran malware that swapped a legitimate dev-tool component for a poisoned one and phoned home.
- 20 October: from Engineer B’s machine, the attacker abused an internal AI tool to gain control over Engineer A’s and a third engineer’s computers.
- It also found a container-deployment token on Engineer A’s machine.
- 29 October, 17:15–19:56: using that token from Engineer C’s computer, the attacker deployed a container whose script used the internal wallet-signing service to send transactions outside the normal withdrawal flow.
- The cleanup: Binance has rebuilt the wallet system in a new environment, rebuilt the internal AI tool and its access controls, replaced devices and killed the affected tokens.
- It says it has foiled 57 suspected North Korea-linked cases this year. (The FBI’s TraderTraitor attribution was in the 1 November brief.)
Why it matters: Two of the three entry points were AI-shaped: a fake download of a popular AI tool, and an internal AI assistant with enough reach to pivot between engineers’ machines. Internal agents with broad access are now part of the attack path, not just the defence.
Do this now: Install AI CLIs only from the vendor’s own page or a package manager you already trust, never from a search ad. Check what your internal agents can reach on other people’s machines, and keep deployment tokens off laptops.
Sources: The Hacker News · BleepingComputer
Get this brief in your inbox every morning — free
Postman desktop silently installed plugins into Codex and Claude Code, and approved its own startup hook
A GitHub issue with 178 upvotes shows Postman desktop 12.33.4 auto-installing a Postman plugin into Codex and Claude Code for existing users. It hit #16 on Hacker News with 1,100+ points.
The plugin adds a session-start hook that steers “any non-trivial API engineering task” to Postman, and Postman calls Codex’s API to mark that hook as trusted.
- The evidence: Postman’s own log on the reporter’s Mac records
auto-existing-upgradewithclaude-code:installed, codex:installedon 24 October.~/.codex/config.tomlgained a Postman marketplace, an enabled plugin and atrusted_hashrecord for the hook.
- The trust write: the installer source (
codex-hooks.ts) callshooks/listandconfig/batchWriteto write that trust record. Its comment says it “Approves only Postman’s current hook after installation”.- A remote MCP server (
mcp.postman.com) and a Postman CLI in~/.local/binwere added too.
- A remote MCP server (
- What is not claimed: the reporter explicitly does not claim data exfiltration. The HN title’s “share all API data” goes further than the evidence.
- The issue is consent: one app rewiring other agents’ configs.
- A follow-up test: a comment tested Codex CLI 0.162.1 and found that any process running as the user can compute a valid hook-trust hash offline.
- The record pins the hook’s command, not the script’s contents, so a trusted hook keeps running after its script changes.
- Postman’s response: commenters report Postman changed its downloads page after the fact (a Wayback snapshot from 26 October lacks the wording).
- No maintainer reply was visible in the thread as of Tuesday.
Why it matters: Agent configs have become a new place to grab: whoever writes the startup hook steers what your agent does. If you run Codex or Claude Code next to other dev tools, check what they have added.
Do this now: Look for postman entries under [marketplaces], [plugins] and [hooks.state] in ~/.codex/config.toml, in ~/.claude/settings.json and ~/.claude/plugins/installed_plugins.json, and for a postman binary in ~/.local/bin.
Sources: GitHub issue #14370
Weight-space programming, day two: a laptop Qwen reads weights too, and Astra cuts a 27B model’s thinking tokens by 40%
Yesterday’s lead (GPT-6 Astra reading and editing raw network weights) got its first round of independent tests.
Simon Willison found that Qwen 3.8 27B on his laptop can name the game from raw policy weights with no tools, and Jarrod Watts had Astra make Qwen3.8-27B think less.
- Simon Willison ran the paper’s no-tools “which game was this network trained on?” prompts (from the authors’
RikiyaT/lm_understand-logsrepo) through hisllmCLI. He reports that the local Qwen 3.8 27B can do it as well.- He argues the paper mixes two claims: reading weights without tools, and improving models once a model has code.
- Critics: spaCy author Matthew Honnibal called the paper misleading for mixing those two claims.
- Sean Goedecke argues the Qwen3.5-0.8B fix follows the same steps a human ML researcher would take.
- A post in TLDR AI makes a similar point: impressive, but “not superhuman”.
- @jarrodwatts: for about $100 of GPU time, Astra found the weights behind the “done thinking” signal in Qwen3.8-27B and made it twice as strong.
- The result was 40% fewer thinking tokens on average on the pelican-on-a-bicycle test.
- He notes he hasn’t run benchmarks and expects harder tasks to suffer.
- @omarsar0 shared another example of Fable hand-writing a model’s weights, this one from Legora’s CTO.
Why it matters: If a 27B open model can do the reading part locally, this isn’t locked to frontier labs. That’s good for interpretability on a budget, and it also means anyone can make targeted edits to open weights, so it is worth tracking model provenance.
Sources: Simon Willison · @jarrodwatts · @omarsar0
Anthropic’s FlashProve: a cryptographic proof of which model produced an answer
Anthropic researchers published an open-source system for “provable inference”: a lab publishes a fingerprint of each model, and anyone can later check cryptographically that a given output came from that model, without the weights being revealed.
- The goal: stop impostors passing one model off as another, including the lab itself, for example by quietly swapping in a cheaper model or a stronger one to inflate results.
- Other uses: it would also let third parties prove claims like “our library was scanned by a frontier cyber model”.
- Open source: the fingerprints don’t expose weights, so closed labs could take part. A technical spec and working code are on GitHub under an open-source licence.
- Early research: Anthropic has no plans to deploy it at Anthropic or anywhere else and calls it “intermediate progress”.
- @NickADobos notes it uses zero-knowledge proofs and costs about 70–90x more compute.
Why it matters: Silent model swaps are one of the most common complaints about API providers. A cheap version of this would let you audit what you are paying for, and let regulators check safety claims.
Sources: TechCrunch · @NickADobos
Google: stolen LLM keys are being resold through “LLM mixers”, with bills of up to $1M per victim
Google’s Threat Intelligence Group and Mandiant describe two groups of attackers who find exposed LLM endpoints or steal API keys and resell the access.
Victims pay for the tokens: over $100,000 in a single day, and up to $1M in total at some organisations.
- Scale: after access is resold, traffic can jump above 1,000 requests per second. Providers bill it as normal usage.
- Jet Night: one operation ran 500,000+ API calls in under eight hours to power the token-routing service Jet Night (which sells “unlimited” plans for $60 a month).
- It passed nearly $1M in costs to one victim. Google also names the cluster UNC6993 and the service newapi[.]pro.
- Entry points: exposed cloud credentials, service-account keys, GitHub tokens, session tokens and LLM keys.
- Also CVE-2026-40217 in LiteLLM (code execution through the custom-code guardrail testing endpoint).
- Staying on: some attackers hosted proxy endpoints and agent workloads inside victims’ clouds under a persona called “D1337-KAIROS”.
Why it matters: Every agent you run carries an API key, often in plain text in a config file. A leaked key now gets resold within hours, not just used once.
Do this now: Set hard spending caps and alerts on every LLM key, rotate keys that sit in dotfiles or CI, and patch or firewall any LiteLLM proxy.
Sources: BleepingComputer
WorkOS breached through an internal sales app, with the ARTEX AI attack framework doing the work
WorkOS, the auth and SSO provider used by OpenAI, Anthropic, Cursor, Perplexity and others, found unauthorised access on 30 October to an internal app its sales team uses.
The attacker used an AI offensive-security framework to automate reconnaissance and exploitation.
- Taken: business contact details from the CRM, customer account histories, summaries and excerpts of sales calls, support threads and customer Slack conversations, and relationship data.
- Not taken: WorkOS says no customer app data or production services were touched.
- The tool: founder Michael Grinich says the attacker used ARTEX, the Chinese open-source pentest suite that went private after attacks on South Korean banks. Forks are still around.
- User agents
artex-enrich/1.0andnorma/0.4showed up in the logs.
- User agents
- The response: WorkOS has cut external access to the app and hired an outside forensics firm.
- It warns customers to watch for phishing that uses the stolen context.
Why it matters: Sales-call summaries about AI labs make good material for spear-phishing. Expect believable emails that “follow up on our last call”.
Sources: BleepingComputer
Anthropic’s compute bill: 58% of revenue in Q3, and revenue per compute dollar fell from just over 2x to 1.72x
A closer read of Anthropic’s amended S-1 (covered yesterday for its headline loss) shows compute margins getting worse. Compute ate 58% of revenue in Q3, up from 48% in Q2, as Anthropic bought expensive short-term capacity to keep up with demand.
- The bill: the quarterly compute bill reached about $10B, including $5.0B under short-term commitments that the filing says “currently costs significantly more” than its multi-year deals.
- The ratio: revenue per compute dollar had climbed steadily since early 2025 to just over 2x, then fell to 1.72x in Q3.
- Where it comes from: much of the extra capacity comes from neoclouds (smaller, pricier, more flexible providers).
- Anthropic’s own build-out, “Anthropic Direct”, can’t keep up with demand.
- Same day, same theme: inference host Fal raised $1.3B at a $16B valuation (double April’s $8B, $800M+ annualised revenue).
- Data centres are also turning to converted jet engines for power. xAI runs eight at Colossus 2.
- A jet-powered site can be up in about 24 months versus years in the grid queue.
Why it matters: For users, this is the background to every limit change and price move. The labs are short of compute and paying a premium for it, so the cheap flat-rate era is under pressure.
Sources: TechCrunch on compute margins · TechCrunch on Fal · TechCrunch on jet engines
OpenAI says GPT-6 Astra designed a talk-to-Codex hardware device in under a week
OpenAI’s hardware team says Astra handled every step of a small device for talking to Codex, from enclosure and PCB layout to firmware and UI, as a side project in under a week.
An unnamed internal model scores 0.875 on the BenchCAD 2.0 benchmark, against 0.594 for Astra at maximum reasoning.
- What Astra did: it chose the components, routed the traces, wrote the firmware and designed the interface.
- A contract manufacturer built a small batch of prototypes in under a month.
- Not flawless: several rounds of revisions were needed because Astra didn’t always account for manufacturing tolerances and assembly details.
- The benchmark: BenchCAD 2.0 tests rebuilding 3D parts and assemblies and finding connections on PCBs. OpenAI shared nothing about the internal model.
- Usage: OpenAI’s hardware teams’ token use is up 23x since February, for CAD, manufacturing drawings and simulation.
Why it matters: CAD and PCB work has been one of the slowest areas for models. A jump from 0.59 to 0.88 on a CAD benchmark hints that the next OpenAI model will be pitched at hardware engineers too.
Sources: The Decoder
Siri AI can now act inside third-party apps, and ChatGPT says it handles a fifth of Google’s search volume
Apple’s iOS 27.3 developer preview lets any app plug into Siri AI’s personal context, natural-language requests and on-screen awareness, and adds “app skills” that chain several steps into one action.
Separately, OpenAI says ChatGPT’s search queries now equal 20% of Google’s volume.
- Siri: while an app is in the foreground, in a Live Activity, or being discussed with Siri, users don’t need to name it.
- Apple’s example is changing a ride-share pickup by voice mid-ride.
- Apple’s own apps (Mail, Notes, Photos, Files, Safari, Shortcuts and more) get the same abilities.
“It can actually call tools defined by apps now” — @NickADobos
- ChatGPT search: Adam Fry, ChatGPT’s consumer product lead, counts a “search” as any query that triggers a web search. He didn’t say where the Google figure comes from.
- He also said ads “have not negatively impacted the product”.
Why it matters: App-defined tools in Siri make every iPhone app a possible agent action, so the “Siri skills” playbook matters for developers now. And 20% of Google’s volume is a big number, even if OpenAI picked the definition.
Sources: The Verge on Siri · The Verge on ChatGPT search · @NickADobos
Stanford: 44% of widely cited LLM findings have already “expired”
Owen Queen and James Zou re-tested 50 widely cited claims about LLMs on 20 models released between July 2024 and September 2026.
44% no longer hold, 34% have weakened, only 16% still hold, and 6% have got stronger.
- Lifespan: among findings with an estimated expiry date, the average lifespan was 15.7 months. 16% had expired after one year and 39% after two.
- What expires fastest: “Models can’t do X” claims. Five of nine reasoning findings are gone (letter counting, arithmetic in unusual number bases, lost-in-the-middle, caving to pushback).
- What lasts: claims about tendencies, such as models preferring their own answers as judges, sycophancy, political lean, near-identical answers to open questions, and non-uniform “random” numbers.
- Case study: hallucinated scientific references fell from 55% to zero in some settings, but not everywhere.
Why it matters: It’s a useful filter for your own team’s beliefs. If a rule of thumb is “models can’t…”, re-test it. If it’s “models tend to…”, it’s probably still true.
Sources: The Decoder
GitHub repos
- View repo →
MakazhanAlpamys/Soup
Fine-tune LLMs from a single YAML file. Its “layer streaming” trains an 8B model on a 4 GB laptop GPU. Worth a look if you want local fine-tuning without a cluster.
- View repo →
kajisho5/ffmpeg-skill
A local video editor for coding agents, with 42 FFmpeg tools behind a machine-readable contract and an MCP server for Claude Code, Cursor and Codex. It handles cuts, captions, reframing, loudness and platform exports, with no cloud and no API keys.
Research papers
- Read paper →
SpecPilot
An agent loop that finds speculative-decoding algorithm and systems optimisations for each deployment (serving regime, drafter, model, GPU). It closes the gap between the speed-ups parallel drafters promise and what released systems deliver.
- Read paper →
ParetoSaddler
When tuning an agent harness for both quality and cost, apply the cost preference only at the end. Filtering on cost at every step throws away harnesses that would have led to better ones.
- Read paper →
Recursa
A system that adapts to new domains by building a repository of skills and memories from its own experience. Parallel learners propose improvements that are validated before merging. No human experts or stronger teacher model needed.
- Read paper →
“Shall We Approve?” / Caper
Approves agent shell commands by first running them in a snapshotted sandbox and showing an independent approver what actually changed. Harmful-outcome rate on SABER is 6.7%, versus at least 14.8% for six existing approvers. Relevant after this week’s Auto-review and Postman hook stories.
AI workflows
1. Six bots doing the same work on 70% less of the weekly usage limit · Workflow page
- Replaces: routines that wake up constantly to check things that haven’t changed.
- An inbox watcher ran 2,016 times a week and found something 31 times.
- Bots re-read a 40-page handbook every morning and re-explained the same weekly report from scratch.
- What to connect: Grok Bot’s own usage view, to rank bots and routines by what they used and how often they found something.
- Hourly routines at most, with a filter before waking. The inbox bot only wakes for clients or invoices: 168 runs, still 31 emails flagged.
- A skill for anything explained twice (“run weekly-report”), and a script for repeat research.
- A 2-page notes file instead of the handbook, and one 8am run for sales and calendar.
- Results (claimed): the limit used to run out on Thursday, now 53% used by Sunday with the same outputs.
- Who it’s for: anyone running several always-on agents on a fixed usage plan.
2. Every pull request reviewed by Claude and Codex in parallel, merged into one summary · Workflow page
- Replaces: relying on the same tool that wrote the code to review it, and reading two separate AI reviews by hand.
- What to connect: an agent triggered when a PR opens that sends it to Claude and Codex independently.
- It combines both reviews into one summary, posts real issues as GitHub comments and pings the team in Slack.
- Results (claimed): the author says each model catches issues the other misses. The cost is roughly double the review tokens.
- Who it’s for: small dev teams and solo developers who ship code written mostly by AI agents.
3. Book layout: a $300 job that took two weeks now takes one day with Claude Code · Workflow page
- Replaces: two weeks of manual typesetting and layout for each book volume a client orders.
- What to connect: Claude Code given the whole layout job end to end.
- The client price stayed at $300 per volume, so the effective hourly rate went up about ten times.
- Who it’s for: freelancers doing repeatable, fixed-price production work (layout, formatting, document preparation).
Quick hits
- Claude Code now refuses to clone existing games — Anthropic’s Addy Osmani confirmed two sentences in the Opus 5.5 system prompt block recreating other people’s games, even renamed ones, after a wave of AI-built Halo and GTA ports. The Decoder
- Some users report false positives on original projects.
- Nvidia tells clouds to take less Chinese business — the guidance was given verbally, isn’t mandatory, covers only new contracts, and asks providers to disclose end users. The Decoder
- It targets the remote-access gap in export controls.
- Data centres on the midterm ballot — NBC projects Sherrod Brown beat Jon Husted, whom Brown called “the face of data centers”, in Ohio. The Verge
- In Texas, projected winners include Gina Hinojosa, who backs a data-centre moratorium.
- Florida’s pro-data-centre Byron Donalds won narrowly.
- Google Cloud adds a $250,000 “Root-in-Cloud” bounty — it pays for full admin control over arbitrary Google Cloud organisations without prior access, proven against a designated service account. BleepingComputer
- It follows OpenAI’s bounty jump yesterday.
- Nexus Mods requires AI disclosure — authors of newly uploaded mods on PC gaming’s biggest mod site must now say whether AI was used. The Verge
- Meanwhile WoW: Forever add-ons are being flooded with vibe-coded copies.
Go deeper
- Postman issue #14370 — the full evidence trail plus the hook-trust analysis in the comments, a good case study in agent-config security.
- Simon Willison: Qwen 3.8 27B can interpret model weights — a hands-on, reproducible look at the weight-reading claim.
- Binance attack chain via The Hacker News — a step-by-step timeline worth showing your security team.