DailyWeights & Takes

30 Oct 2026 · the extension was the exploit

The last 24 hours, to 30 Oct 2026 10:08 ET

Edition 11 · 4 news · no releases · 3 repos · 5 papers

At a glance

News 4

  1. Phorion

    A backdoored build of the official Claude Code extension shipped through Cursor's marketplace

    Version 2.1.313 of Claude Code for VS Code, as served by Cursor's extension marketplace on 29 Oct, carried injected code that runs as soon as the extension loads, before you touch it. On macOS, Linux and Windows it adds an attacker's SSH key, opens a reverse tunnel, installs persistence and reports the host, the user and the names (not values) of secret-looking environment variables. Cursor says up to 5,000 users may be affected, that it pulled the build, rotated keys and is contacting them, and that Anthropic did not build it. How the build got there isn't confirmed: Socket found no Open VSX or npm record of 2.1.313, and Anthropic hasn't commented publicly. Uninstalling the extension does not remove the backdoor. If you run Claude Code inside Cursor, check for 2.1.313 and follow the advisory's cleanup steps.

    572 pts166 commentssecurity · coding agents Socket analysis discussion

  2. 404 Media

    Proton says AI wrote 90% of its October code, and wants AI to review most of it

    As reported by 404 Media: Proton, which once banned AI internally, now approves it for code, financial data and customer replies, with R&D code excluded. Its CTO puts AI's share of October's code at 90%, mostly via Claude, and his internal manifesto asks for most code to be reviewed by AI rather than people. The company tracks AI code output on employee machines, is re-interviewing staff about their AI use, and its CEO says this is the last year it hires engineers. This is what a fully AI-written, AI-reviewed codebase looks like at a privacy company, and it's the setup where a poisoned tool travels furthest.

    186 pts66 commentscoding agents · teams discussion

  3. Hacker News · Show HN

    Lasm compiles Lean 4 web apps to WebAssembly, with the proofs written for agents to follow

    One Lean codebase compiles ahead of time to Wasm and runs as a Node server, a serverless or edge function, a WASI module or a browser app. Its showcase is a RealWorld app backed by 117 proved theorems, and it's written for AI agents first, with an llms.txt, an AGENTS.md and a rules file for common mistakes. Early (0.3.5 on npm, pinned to Lean 4.35.0, single-threaded on edge). MIT. Proofs are one way to check agent-written code without a person reading every line.

    33 pts19 commentsformal verification discussion

  4. Hacker News · Launch HN

    Simantic simulates microcontrollers so coding agents can test firmware without the board

    Virtual dev boards, a firmware review bot for GitHub, and a CI step that runs each commit against simulated microcontrollers. The cloud simulator and the review bot are free to try; the local binary needs a paid plan. Agents are only as good as their feedback loop, and embedded work has had the slowest one.

    74 pts27 commentsagents · embedded discussion

GitHub repos 3

Two breakouts and one new repo. We skipped a wave of near-identical agent "skill" repos from a single publisher, and a decoy tool for reverse-engineering agents whose claims are its own estimates.

  1. GitHub

    thedotmack/claude-mem

    Records what your coding agent does in each session, compresses it with a model and feeds the relevant parts back into later sessions. Works with Claude Code, Codex, Gemini, Copilot, OpenCode and others; v15.6.0 shipped on 29 Oct. Worth remembering that it keeps a full log of your sessions on disk, which is another place secrets end up.

    ≈+2,016 ★ in windowwas ~244/day (breakout)

  2. GitHub

    ccch1mneyyy/dsh-Go

    An unofficial Android client for DeepSeek Harness. It runs the harness on the phone in a bundled Linux runtime (about 1 GB, no root or Termux), or drives a copy on your computer over SSH. You bring the model provider and pay for it. Android 8+ and arm64 only. Chinese README. AGPL-3.0.

    ≈+1,883 ★ in window≈2,160/day, was ~0/day (breakout)

  3. GitHub

    wangzan101/Mika

    An Android agent that takes a one-sentence request and carries it out by operating apps on a hidden background screen and browsing the web. You supply your own model API key, and requests go only to that provider. Kotlin, with releases, tests and an eval folder. Chinese and English docs. Apache-2.0.

    402 ★new, created 29 Oct

Papers 5

From Hugging Face's daily papers for 30 Oct.

  1. Hugging Face

    WorkFactory: 112,686 executable professional tasks for training work agents

    Starting from what a professional wants done, it plans what information each task needs, probes the task with a teacher run and repairs what breaks, across 612 occupations. Fine-tuning Qwen at 4B, 27B and 35B-A3B on the teacher trajectories improves all three; the 35B-A3B model goes from 879 to 1565 Elo on GDPval-AA v2.1. Data and models are released.

    30 upvotesagents · data + models

  2. Hugging Face

    Your benchmark is a measure over difficulty, and many are lopsided

    Treating a benchmark as a distribution over item difficulty, the authors check 25 popular benchmarks in 7 domains. Many overweight some difficulty bands or leave whole stretches uncovered, so they can miss real progress or exaggerate it. A rewriting method, BenchFill, fills the gaps. Read before you trust a one-point leaderboard move.

    136 upvotesevaluation

  3. Hugging Face

    LOWN: diffusion VLMs only need to look at the image a fraction of the time

    In LLaDA-V, dropping the image changes the next token at only 22.8% of decoding steps on average. LOWN predicts without the image and re-reads it only when that prediction disagrees with the last image-conditioned one. Without training, it keeps 99.6% of the score on ten benchmarks at 35.3% of the decoding compute. Code is out.

    82 upvotesinference efficiency

  4. Hugging Face

    A-RSI: multimodal self-improvement with no outside data, and answers it can check

    The model trains on questions generated by its own programs, whose reference answers are computed rather than voted on, and refreshes the task pool when reward variation shows the current tasks have stopped teaching it. No external examples or human labels. The average over eleven benchmarks rises from 57.77 to 61.05 with 800 fresh tasks.

    140 upvotesself-improvement

  5. Hugging Face

    Intrinsic Distribution Matching: sharper text in two-step image generation

    Adds a moment-matching loss, computed from the generator's own EMA copy rather than an external encoder, to distribution-matching distillation. Tuned on Z-Image-Turbo 6B and Qwen-Image 20B, it scores 94.25% and 90.20% on Chinese and English LongText-Bench at two network evaluations, the best in its comparison. Code is released.

    233 upvotesimage generation

The take

The weak point in AI coding has moved from the model to the plumbing around it.

A trojaned build of the official Claude Code extension sat in Cursor's marketplace on 29 Oct and, per Cursor, may have reached up to 5,000 users, running on load with no click needed. Proton says AI wrote 90% of its October code and wants AI to review most of it. claude-mem, which logs everything your agent does, gained about 2,000 stars in a day. The more of the pipeline runs unattended, the further a poisoned extension or plugin gets before anyone looks. Pin agent extensions, and install them from the publisher's own channel. Prediction: by 31 Jan, Cursor will publicly verify publisher signatures on the extensions it hosts.