A primary-source guide to indirect prompt injection in 2026: Google's 32% increase, Unit 42's 22 techniques, Comet browser demos, and the folk stats to ignore.
A practical prompt injection threat catalog for SMBs letting Claude Code or other AI agents browse the web, with safe research defenses.
Google Lighthouse is a powerful lab diagnostic, not a real-user verdict. Here is what it actually measures, why scores swing between runs, how the zombie "20% per second" stat overstates real research, and what credible performance data looks like.
Anthropic's Claude Security launched in public beta April 30, 2026. Unlike traditional SAST tools, it uses LLM reasoning to trace data flows across multiple files and find complex vulnerabilities. Key results: 500+ bugs found in production open-source code, 79% false positive reduction, 1,870 analyst-hours saved in 30 days. Integrates with CrowdStrike, Palo Alto Networks, Microsoft Security, SentinelOne, TrendAI, and Wiz.
Anthropic's internal security team built CLUE (Claude Looks Up Evidence), an agentic AI detection and response platform that cut false positives from 33% to 7% and saved 1,870 analyst-hours in 30 days. Here's how they built it and what security teams can learn from it.
A step-by-step install guide for the Safe Web Research skill — TypeScript hooks that intercept and sanitize every web call Claude makes, plus judgment rules that tell Claude what to do when something looks suspicious.
A walkthrough of the /safe-web-research playbook: how the sanitiser hook wraps every web fetch, what risk signals trigger an abort, and how the skill plans a safe AI search path around tarpits, cloaking, prompt injection, and infinite-maze trap doors.
AI tarpits like Nepenthes, Iocaine, and Cloudflare's AI Labyrinth trap unwanted scrapers in infinite mazes, while Anubis blocks them with proof-of-work. The catch: blanket-blocking AI also nukes the retrieval bots that surface your site in ChatGPT, Perplexity, and Google AI Overviews. The fix is per-bot policy, not all-or-nothing.
AI tarpits close on behavioral patterns, not identities. The Resilient AI Web Research Protocol maps all seven tarpit trigger signals to behavioral rules so an agent following the protocol is never the crawler these traps are set for.
