Anthropic Releases Claude Mythos Preview Vulnerability Report

#1 𝕏

Anthropic published a detailed technical report on software vulnerabilities and exploits uncovered in Claude Mythos Preview, outlining the specific flaws, attack vectors, and mitigation strategies.

Also covered by: @Anthropic, @Boris Cherny, @Greg Isenberg, @Simon Willison

#2 𝕏

Mustafa Suleyman praises Jordi Rib1 and the Bing team for rapidly shipping the Harrier open-source embedding model across Microsoft AI with outstanding speed and quality—details in today’s blog post.

#3 📝 Simon Willison

GLM-5.1: Towards Long-Horizon Tasks - Chinese AI lab Z.ai released GLM-5.1, a 754B-parameter MIT-licensed model available via OpenRouter; Simon used it to generate an excellent SVG pelican but encountered broken CSS animations which the model helped diagnose and fix, and later produced a possum-on-an-escooter variation.

#4 𝕏

Dario Amodei is granting defenders early, controlled access to Mythos Preview so they can uncover and patch vulnerabilities before Mythos-class models proliferate.

Also covered by: @Anthropic, @Boris Cherny, @Greg Isenberg, @Simon Willison

#5 𝕏

Kevin Weil 🇺🇸 launched Prism’s new Paper Review AI workflow for technically vetting scientific papers, boosting rigor, correctness, and reproducibility.

#6 𝕏

LlamaIndex 🦙 teamed up with LanceDB to launch a structure-aware PDF QA pipeline using LiteParse for structured text and screenshots, Gemini 2 embeddings in LanceDB, and a Claude agent for text+image reasoning—achieving near-perfect accuracy across most tasks.

#7 📝 Anthropic Engineering

Quantifying infrastructure noise in agentic coding evals - An investigation showing that infrastructure configuration can materially affect agentic coding benchmark results, sometimes changing scores by several percentage points—more than differences between top models. The piece emphasizes the importance of controlling infra variables when evaluating agentic systems.

#8 𝕏

Harrison Chase announced that LangSmith Fleet now integrates with Arcade.dev, offering enterprise-grade access to 8,000+ tools and enabling you to build no-code Claude Cowork/OpenClaw–style agents in minutes.

#9 𝕏

Harrison Chase unveils LangSmith’s tracing and evaluation platform—spotlighted on new SF & NYC billboards—to help teams track, diagnose, and optimize agent behavior in real-world conditions.

#10 𝕏

Santiago built a 100% open-source CLI that maps your database schema to context files, letting you spin up a full AI data analyst in under two hours—entirely local, no SaaS or API keys required.

#11 in

Udi Menkes highlights Andrej Karpathy’s new RAG-replacing architecture: a three-layer, LLM-driven markdown wiki (raw sources, AI-maintained knowledge pages, schema) that compounds knowledge across 10–15 pages per input.

#12 in

Romain Huet live-demoed OpenAI’s Codex app—building a small game with Peter Yang and Alexander Embiricos—and underscored that as AI agents multiply, blurred roles (designers coding, engineers owning product) demand stronger judgment, taste, and user-centric insight.

#13 𝕏

Mustafa Suleyman launched Harrier, an upgrade to Bing’s web grounding that uses enhanced embeddings for more accurate retrieval, stable agent behavior across multi-step tasks, and support for 100+ languages.

#14 📝 Jesse Vincent

Rules and Gates - The author describes discovering the concept of a "gate" in prompts while building Superpowers, a term credited to Claude Code. The post previews a discussion of how gates influence prompting and agent behavior.

#15 𝕏

Santiago showcases @modulate_ai’s Velma deepfake detection model hitting 98.9% accuracy on Hugging Face’s Arena at 120× lower cost, making continuous, real-time voice-call screening possible to combat deepfakes.

#16 𝕏

Cursor introduced Design Mode in Cursor 3, enabling users to annotate and precisely target browser UI elements.

#17 𝕏

Lenny Rachitsky breaks down how Anthropic uses Claude to automate lead outreach, email sequence generation, A/B testing and performance analytics. This has accelerated campaign velocity and let the team scale acquisition without adding headcount.

#18 𝕏

DeepLearning.AI spotlights Andrew Ng’s insight that rapidly improving voice-based AI interfaces will enable more natural, accessible interactions alongside traditional UIs.

#19 in

Peter Yang wires Google Workspace, Mercury and other APIs into his OpenClaw AI agent to automate the first 80% of docs, slides and analytics before he polishes the rest.

#20 in

Anu Jagga Narang highlights that benefits capture our hoped-for gains but only impact reveals real outcomes—features may hit north-star metrics yet burn out support or break other teams, so PMs must measure both to tell the full story.

#21 in

Marc Baselga shows how an Adobe product lead with zero coding skills set up Claude Code to turn a folder of markdown files into an AI chief of staff.

#22 𝕏

claire vo 🖤 shows that one PM now managing the output of 20 engineers means any mediocre PMs you had before are simply mediocre at 5× scale.

#23 𝕏

Dario Amodei says Glasswing is just the first step in a multi-year effort to patch and secure the world’s software infrastructure. He emphasizes this will demand unprecedented collaboration among AI companies, cyberdefenders, software providers, governments, and more.

#24 𝕏

Kevin Weil 🇺🇸 built a structured Codex skill on GPT-4 (5.4 Pro), adding more scaffolding to further improve its already strong performance.

#25 in

Kevin Weil launched Paper Review in Prism, an AI-powered workflow that scrutinizes math, derivations, notation, units, structure and claim support in scientific papers.