Unsupervised Learning NO. 526

Hey! Hope you all are doing well!

PAI 5.0 is imminent, and I rebuilt the homepage for the project using PAI’s integration of Claude Design. Coolest part? Claude Design is a GUI application, not an MCP or API. But it was still completely automated using my buddy Ron’s Interceptor tool!

GET INTERCEPTOR
OURPAI.AI

Couple of blogs.

Pliny the Liberator used Opus to jailbreak Opus 4.7. THE THREAD

Vercel tells customers about an April 2026 access incident It seems likely that this will be a pretty big one, given how many people use Vercel.

VERCEL SECURITY BULLETIN | MALLORY REPORT

Researcher finds multiple vulnerabilities in Lovable Supabase Implementation Lovable-hosted “vibe-coded” apps can come with broken Supabase row-level security, and one researcher found critical flaws exposing 18K users’ data. The Register reports Lovable’s security scan exists but the app owner didn’t implement fixes. THE REGISTER ARTICLE

LLM-tier personal computer security should be your next home hardening plan This LessWrong post argues AI will make supply-chain attacks and phishing easier, so you should sandbox, firewall, and use phishing-resistant keys for critical accounts. IMPROVING YOUR PERSONAL COMPUTING SECURITY TO DEFEND AGAINST AI

Nearly 6 million internet-facing FTP servers still exposed. ARTICLE

AI misinformation becomes SEO “consensus” and traps users in loops Lily Ray shows how fake “core update” claims get scraped, cited, and repeated by AI overviews until they feel real. THE AI SLOP LOOP

Anthropic’s MCP defaults let attackers execute OS commands via RCE Cybersecurity researchers say Anthropic’s Model Context Protocol SDK has unsafe STDIO defaults that allow arbitrary command execution across thousands of servers. THE HACKER NEWS ARTICLE

NSA quietly uses Anthropic’s Mythos despite Pentagon’s “supply chain risk” fight Axios reports the NSA is using Mythos even as DoD tries to block Anthropic and argues in court it endangers national security. AXIOS SOURCE

Claude Opus system prompt changes from 4.6 to 4.7 Simon Willison walks through Anthropic’s published system prompts, comparing Opus 4.6 vs 4.7 and highlighting safer child handling, less pushiness, and tool-search behavior. SIMON WILLISON WEBLOG

Anthropic’s $800B valuation offers tied to $30B run-rate Anthropic reportedly got investor offers valuing it around $800 billion, after revenue surged to about $30 billion annualized. THE NEXT WEB ARTICLE

Looks like OpenAI models don’t work well with OpenClaw TWITTER THREAD

AI use is rising at work, but adoption is messy ZDNET reports Gallup data: half of US employees use AI, yet many don’t know company strategy, workflows shift unevenly, and friction wastes hours. AI WORKPLACE CHANGES

Salesforce wants chat-and-agents to replace the developer UI Salesforce says Headless 360 exposes its platform via APIs and MCP tools, so humans and coding agents can compose apps anywhere. It’s also rolling out Agentforce Vibe 2.0, testing, and observability.

This is Company as API. It’s happening! SALESFORCE BETS ON CONVERSATION AS THE NEW INTERFACE FOR DEVELOPERS

Netgear gets an FCC okay to keep foreign-made routers Quartz says the FCC conditionally lets Netgear sell new models and update existing ones until 2027, while rivals face update cutoffs. QUARTZ ROUTER EXEMPTION

Cloudflare turns AI into a unified inference layer for agents Cloudflare says AI Gateway gives you one API to call models from many providers, manage costs, fail over automatically, and stream resiliently. CLOUDFLARE AI PLATFORM

Cloudflare introduces an Agent Readiness score for sites Cloudflare launches isitagentready.com to score how easily AI agents can discover, read, access, and use your site. It also adds the checks to URL Scanner. AGENT READINESS TOOL | CLOUDFLARE URL SCANNER

No One You Love Is Ever Dead, in Hemingway’s grief letter Maria Popova shares Hemingway’s March 19, 1935 letter after a son’s death, using it to argue that love outlasts death. THE MARGINALIAN ARTICLE

AI can mimic consciousness but can’t instantiate experience ABSTRACTION FALLACY

Claude can stylometrically identify you from your writing, scary Patrick Stevens shows an “incognito” Claude still guesses his name by matching his unpublished style, arguing anonymity is basically over. CLAUDE KNOWS YOU

AI’s token scarcity is creating a real intelligence class divide Nilesh Jasani argues platforms are throttling, trimming depth, and shifting costs from “free access” to “buy compute,” locking advantages into moats. TOKEN INEQUALITY AI HAVES AND AI HAVE-NOTS

Consumption choices might shape your identity more than your job does Noah argues that producing gets status, but consuming forces real self-questioning. He thinks AI could make life feel like college again, if we share the gains. CONSUME-TO-IDENTITY ESSAY

Boredom keeps your mind alive while algorithms steal the space Michael Pollan argues boredom lets spontaneous thought happen, while phone scrolling and chatbots outsource consciousness, dulling you. https://nautil.us/defending-our-consciousness-against-the-algorithms-1279260/ | THE RELATED TOPICS SITE

AI turns Dunning-Kruger into active sycophancy AI AND HUMAN COMPETENCE

Magika uses a tiny deep model to detect file types fast Magika is an AI file type detector that runs a small deep learning model to classify files quickly. It ships as a Rust CLI plus Python and other bindings, with confidence modes and JSON/JSONL outputs. GOOGLE MAGIKA README | GOOGLE MAGIKA CLI PACKAGE

Voicebox gives you local voice cloning, effects, and an API Voicebox is a local-first voice synthesis studio that clones voices, generates speech across multiple engines, and applies effects in-app. It also exposes a REST API so you can embed voice generation into your own projects. GITHUB REPO | LOCAL-FIRST VOICE CLONING | MULTI-ENGINE TTS STUDIO | REST API FOR SELF-HOSTING

ikno turns your actual work logs into instant daily recaps It grabs what you already did from git, notes, and Claude Code sessions, then uses an LLM to write a recap in your chosen style. This is the type of thing that our DA will be managing for us. GET STARTED

Claude-Code-Glow adds a peripheral signal for Claude Code prompts. PROJECT TINY MAC SETUP

The World Leaks the Future: Harness Evolution for Future Prediction Agents ARXIV PAPER

Make absolutely sure that anything you are building with AI and putting online is properly secured (or at least not a complete shitshow).

I have a whole entire security system designed to continuously audit all the stuff I have out there. And I’m improving it constantly. I recommend everyone does the same.