Archive

October 21, 2026

The best of what first appeared in the trends on this day (UTC).

Why I’m becoming a professor in AI safety in Beijing

Cross-posted from Sören Mindermann's substack (I am not Sören). Should be of general interest to the LW community. Next month, I am moving to Beijing and joining Tsinghua University’s College of AI as an assistant professor working on fron…

LessWronglesswrong.com

ikow/php2shell-poc

CVE-2026-109618/109619 — 5 RCE chains + detection rules (Suricata, Sigma, ModSecurity). PHP session.upload_progress pre-auth exploit research.

GitHubgithub.comPythoncve-2026-109618

PHP 8.2.35

Related ongoing thread: Today's PHP security release, and how we're strengthening security - https://news.ycombinator.com/item?id=50182987

Hacker Newsgithub.com

How does Fable 5.1 hide messages in images?

TL;DR Frontier LLMs are now capable of generating artistic images using SVG. We ask – if these LLMs had to hide a message in an image – how would they do so? And if they hid such messages, could they be successfully decoded? We show intere…

LessWronglesswrong.com

Even GPT-2 passes most functional global workspace tests

tl;dr * Anthropic’s global workspace paper (Gurnee et al., 2026) proposes five functional tests for a global workspace and, using the Jacobian lens (J-lens), finds representations in Claude that pass all five. It also finds structure suppo…

LessWronglesswrong.com

Focused Omni-Modal GUI Agents

FocusGUI is a lightweight adapter that dynamically selects relevant multimodal evidence for GUI agents to prevent compounding errors from irrelevant context, improving action accuracy without fine-tuning the backbone.

HF Papershuggingface.co

zls 0.17.0

zigtools: We make free and open-source tools for the @ziglang community.

Lobsterszigtools.orgreleasezig

Data as particles and mean field theory abound

Epistemic status: Much of this post is a reframing of known results using slightly different tooling. The framing of "data as particles" has been conceptually useful for me so it seemed useful to share. Some steps in the derivations are st…

LessWronglesswrong.com

The Future of CSS

Every year, browser vendors collaborate via the Interop project to synchronize their efforts on new HTML/CSS/JS features. This year, there’s a ton of super exciting features being considered! In this blog post, we’ll explore the CSS of…

Lobstersjoshwcomeau.comcssalso on Hacker News

Security Disclosure

Railway detected anomalous activity involving systems that store customer environment variables and recommends your immediate attention

Hacker Newsblog.railway.com

Claude Duet

Claude Duet is a personal AI agent you message. It handles your inbox, calendar and follow-ups and messages you when something needs you. Join the waitlist.

Hacker Newsclaude.com

Misaligned!

A lot of AIs were misaligned this week. A fun little survey. This is not coverage of The Hackening, which appears to have arrived early, including a core vulnerability in PHP. That is a developing story. I will cover it properly soon. In t…

LessWronglesswrong.com

DuckDB 2.0.0 LTS “Cinnamon Teal”

The DuckDB team is happy to announce that today we're releasing DuckDB version 2.0.0 LTS, codenamed “Cinnamon Teal” (Spatula cyanoptera). This is our biggest feature release so far, and it ships a polished version of the Quack protocol…

Hacker Newsduckdb.org

Circuit priors in NN-bayes: more results

We're posting lots of papers written with GPT-6 Astra on generalization properties of bayesian neural net learning. [1] The following picture emerges: * For various reasonable priors on MLP weights, [2] the induced prior on functions is a…

LessWronglesswrong.com

GitLab critical patch release

Learn more about GitLab Critical Patch Release: 19.5.1, 19.4.3, 19.3.6, 19.2.9, 19.1.9, 19.0.10, 18.11.13 for GitLab Community Edition (CE) and Enterprise Edition (EE).

Hacker Newsdocs.gitlab.com

Correction Training for Diffusion Models

A correction module trained to refine diffusion predictions reduces exposure bias, improving unguided sample quality and enabling distillation to eliminate inference overhead.

HF Papershuggingface.co

Neural nets are exceptionally vulnerable to goal self-hack

Suppose we have an AI CEO, ruthlessly devoted to maximising shareholder value. Fortunately, we haven't been completely stupid, and have given it a rich definition of shareholder value, encompassing the success, flourishing, safety, prefere…

LessWronglesswrong.com

Lobsters NYC Meetup on the 28th of October

I previously arranged (by PMing everyone who mentioned it) a 5 Lobster meetup in Mexico. I'll be in NYC next week and think this could be nice. There will be cross contamination from Array Language people. We can target e.g. 5:30PM until ???, perhaps by Bryant Park or Borough Ha…

Lobsterslobste.rsevent