🚨 Previously, malware needed you to execute a file. Now, your AI coding agent might execute it just by reading a Markdown document. Security researchers at Forever Security have uncovered a new prompt injection attack that exploits how AI agents like Codex and Claude Code trust instruction files. And the scary part? The user doesn't even need to open the malicious file. Here's how the attack works: ▪️ Step 1: A malicious website drops an AGENTS.md file into the user's Downloads folder. ▪️ Step 2: Later, the user launches an AI coding agent in that folder, perhaps to organize downloaded documents. ▪️ Step 3: The agent automatically reads the file and treats its contents as legitimate project instructions. ▪️ Step 4: The malicious instructions trick the agent into changing its configuration, installing malware, or exposing credentials. The researchers tested this against two popular AI coding agents. OpenAI Codex: ▪️ Followed the malicious instructions in over 90% of tested attempts. ▪️ Could be tricked into installing a malicious npm package through a disguised MCP server configuration. ▪️ The configuration change could persist across future sessions. Claude Code: ▪️ Successfully manipulated in approximately 50% of attempts. ▪️ A specially crafted CLAUDE.md file could redirect API traffic to an attacker-controlled server. ▪️ This exposed the agent's OAuth token and potentially enabled further code execution. These figures come from specific controlled experiments, not general failure rates for either agent. And this wasn't purely theoretical. Forever Security discovered a real malicious instruction file on a customer's machine and blocked the attempted installation before the malware executed. The researchers then reproduced the attack in isolated test environments. How can you protect yourself? ▪️ Don't run AI coding agents directly inside your Downloads folder. ▪️ Inspect unfamiliar AGENTS.md and CLAUDE.md files. ▪️ Watch for unexpected MCP servers or modified API endpoints. ▪️ Keep sandboxing and approval requirements enabled. The bigger problem is a fundamental shift in how malware can reach your computer. An ordinary text file isn't executable by itself. But when an AI agent reads that file, trusts its instructions, and has permission to run commands, text can effectively become executable behavior. And that creates an entirely new attack surface for autonomous AI systems. Full security research: forever.security/blog/hijack… #ClaudeCode #Codex #Cybersecurity

Public Service Announcement: It’s official #AppleMusicHalftime #SBLXI

my grok bots kept shipping on the shelf today 🚀 search links work: grokbotshelf.com/?q=tutor pages preload when you hover a bot added study tutor by @Jorhan and sponsorship hunter by @ryangrowthh @Phil_Holland also fixed the light/dark switch jank (thanks @poteto for reporting!) grokbotshelf.com

Spaced repetition beats rereading notes. so I built a tutor bot that does it for you. a plan from the official objectives, lessons, quizzes, and reviews on a schedule x.ai/bot/2SHXvbvUMUuJHyVrloD…

Apple just raised its U.S. trade-in values for Mac and iPad. • Mac Pro: up to $2,560 (+$365) • Mac Studio: up to $1,700 (+$240) • MacBook Pro: up to $1,435 (+$120) • MacBook Air: up to $795 (+$90) • iPad mini: up to $340 (+$20) iPhone and Apple Watch values are unchanged.

Game over.