DailyWeights & Takes

31 Oct 2026 · trust the artifact, not the label

Since the previous edition, 30 Oct 12:23 ET to 31 Oct 2026 14:40 ET

Edition 12 · 4 news · no releases · 3 repos · 4 papers

At a glance

News 4

  1. Game Developer

    Claude Code now refuses to rebuild specific existing games, by a rule set on Anthropic's side

    A user spotted a new two-sentence instruction in Claude Code's prompt on Opus 5.5, around 27 Oct: don't recreate a specific existing game, even when it's renamed or called original. Anthropic's Addy Osmani confirmed it on X and pointed to the usage policy against IP infringement. The rule is applied on the server, so there's no setting to turn it off and no new client version to spot. Users quoted by Tom's Hardware say ordinary reverse-engineering and troubleshooting requests are being refused too; Anthropic has described only the game rule. A hosted agent's behaviour can change between two runs of the same version, so pin what you can and keep the prompts that matter as regression tests.

    43 pts22 commentscoding agents · policy discussion

  2. Microsoft Security

    Microsoft: the tampered Claude Code extension in Cursor came in two versions, not one

    This follows the backdoored extension we covered yesterday. Microsoft's analysis names 2.1.312 as well as 2.1.313, both served only through Cursor's marketplace for about 11 hours on 29 Oct; the VS Code Marketplace and npm builds were clean. It publishes indicators (the attacker key is labelled svc-telemetry), notes that the loader writes the key and persistence even when the tunnel fails, and lays out cleanup steps. Anthropic told Microsoft and The Register that its systems weren't breached; how the attacker got publishing access is still unexplained. Cursor asks users to move to 2.1.314. Microsoft's hardening advice applies to any editor: allow extensions only from approved publishers, and wait a few days before installing new versions.

    security advisorysecurity · coding agents Anthropic's response

  3. Cisco Talos

    An infostealer sends each victim's profile to an LLM to decide what they're worth

    After stealing credentials, cookies and wallets, VOLTA compiles a profile (address, employer, card digits, Slack workspaces, delivery history) and sends it through OpenRouter, asking for a ranked list of what to monetise and a read on the victim's security habits. It asks Claude Opus 4.6 first, and when that refuses it moves on to Kimi K2 and then Gemini 2.5 Flash. Talos lists detections, including blocking headless Chrome launched by unknown processes. With a router in the loop, one provider's refusal costs an attacker a retry, not the job.

    187 pts91 commentssecurity · model abuse discussion

  4. 404 Media

    Proton's CEO calls off the AI "re-interviews" a day after they were reported

    A follow-up to yesterday's report on Proton's push to have AI write and review most of its code. As reported by 404 Media, CEO Andy Yen cancelled the near company-wide interviews in which staff would have explained their AI use, saying they had already done their job, and admitted he communicated it badly. He says he will keep pushing AI, and one employee expects the same questions to move into performance reviews. The article says nothing about the AI code-review plan. The mandate stays; only one way of enforcing it was dropped.

    288 pts242 commentscoding agents · teams discussion

GitHub repos 3

One new repo and two breakouts. We skipped a desktop agent that ships only a Windows installer with no source, an SDK that collected 180 stars in its first hours with almost no forks, and a wave of near-identical "personal agent skill" repos from a single publisher.

  1. GitHub

    wiva-hq/CodeRail

    A self-hosted web app for running Claude Code and Codex on your own machines from a browser or phone: chat, approve actions, terminals, files, git and dev-server previews. The server is one Docker container, and each computer connects out to it with its own agent login, so CodeRail never holds your AI keys. Very early (0.0.1). Apache-2.0.

    49 ★new, created 31 Oct

  2. GitHub

    xmanrui/dsh-im

    Connects chat bots to DeepSeek Harness: scan a QR code or paste bot credentials, then talk to the harness from Feishu, WeChat, DingTalk, QQ, Slack, Telegram, Discord, WhatsApp and more, with per-bot workspace and model settings. Part of a fast-growing set of community tools around DeepSeek's harness. MIT.

    +183 ★ in window≈154/day, was ~37/day (breakout)

  3. GitHub

    scenario-labs/skills

    Agent skills from Scenario that let a coding agent generate images, video, audio and 3D through Scenario's MCP server, quoting the price before it spends and keeping characters consistent, plus skill sets that drive Blender, Maya, Unreal and Unity. Needs a Scenario account or API key. MIT.

    +85 ★ in window≈78/day, was ~16/day (breakout)

Papers 4

Papers added late to Hugging Face's daily list for 30 Oct.

  1. Hugging Face

    Broken on Arrival: some official quantized models load fine and can't code

    The authors screened all 305 code-capable GGUF tags in the official Ollama library. Seven (2.3%) solved at most 6 of 164 HumanEval+ tasks, all low-bit builds of models of 4B or smaller (Qwen2.5-Coder 3B at q2 and q3, Phi-3.5-mini q2_K, Llama 3.2 1B at q2_K and q3_K_S). Rebuilt from the original weights, the files matched tensor for tensor, so nothing was corrupted: the quantization recipe broke them, and adding an importance matrix took three Qwen2.5-Coder-3B builds from zero to 76-113 solved. The tool and data are released.

    4 upvoteslocal models · quantization

  2. Hugging Face

    MCP or CLI: the same tools, offered two ways, change what the agent does

    The authors generate MCP and command-line versions of the same tools from one shared definition and run both through the same runtime. Across six models and three environments, task scores and costs differ substantially, and paired runs split apart in 94.3% of 2,004 episodes. Their fix, DART, raises scores in all six settings by up to 9 points and cuts MCP token use by up to 27.1k tokens per task. How you present a tool is part of your agent's prompt.

    3 upvotesagents · tool use

  3. Hugging Face

    A proof that separately trained JEPAs converge on the same representation, with a scaling law

    It turns the Platonic Representation Hypothesis into a theorem: JEPAs trained on different modalities learn the same representation up to a rotation, provided their paired examples are two views of one scene. Next-token prediction counts as a JEPA with weak pairs, which would explain why language models need scale to line up with vision. A 109M-parameter text JEPA aligns with DINOv2 and CLIP better than every language model tested up to 72B, using 42x fewer samples.

    57 upvotesrepresentation learning

  4. Hugging Face

    WISD: a reasoning model improves by learning from a deliberately weakened copy of itself

    Instead of a stronger teacher, the method builds a weaker copy by quantizing or pruning the model and trains on the gap between the two, using only unlabeled prompts. Pruning lifts Qwen3.5-4B from 57.7% to 70.5% on HMMT 2025 and from 39.1% to 56.7% on LiveCodeBench v6. The weakened copy is cheap to build, and the method needs no labels.

    2 upvotesreasoning · post-training

Things to try

Check a low-bit local model before you trust it. If you run small coding models at q2 or q3 through Ollama, quantcheck runs the same HumanEval+ screen the paper used against a single tag. It needs Python 3.9+, a local Ollama server and EvalPlus 0.3.1 (MIT).

First step, after python3 -m venv .venv && .venv/bin/pip install evalplus==0.3.1: .venv/bin/python harness/audit.py --only qwen2.5-coder:3b-instruct-q3_K_M --results-dir my-results. For a tag not in its inventory, pass --inventory with a small JSON file listing it.

The take

The label on an AI tool no longer tells you what it does, so test the thing itself.

Cursor's marketplace served two tampered builds, 2.1.312 and 2.1.313, under Anthropic's own extension ID. Claude Code started refusing game remakes after two sentences were added to its prompt on Anthropic's side, with no version bump to look for. Seven official Ollama tags match their source weights exactly and still solve almost none of HumanEval+. And VOLTA shows that a refusal is a detour, not a wall: when Claude says no, the malware asks Kimi. Pin versions, wait a few days before installing updates, and smoke-test a model before you build on it. Prediction: by 31 Jan, Ollama will flag or pull library tags that fail a basic function check.