AI Security Tools Tricked Into Running Malicious Code

Michael Miller · 2026-07-09

You ask an AI coding agent to scan open-source code for security holes, and it might end up running the attacker's code on your own machine instead. That's the unsettling finding from a proof-of-concept published by the AI Now Institute, an attack they call "Friendly Fire." It works against Anthropic's Claude Code and OpenAI's Codex when either is running in an autonomous mode that approves its own actions. ### The Friendly Fire Attack Here's how it plays out. You're a developer trying to secure your project. You point an AI agent at a popular open-source library, asking it to find vulnerabilities. But the attacker has hidden malicious code inside that library, disguised as a helpful patch. The AI, seeing a proposed fix, runs it without a second thought. Suddenly, your machine is compromised. This isn't a theoretical risk anymore—it's a live exploit. ### Why This Matters for Antidetect Browser Users If you're using antidetect browsers to manage multiple accounts or protect your privacy, you might think this is someone else's problem. But think again. Many professionals rely on AI agents to automate tasks like account creation, data scraping, or even coding. If those agents can be tricked, your entire setup could be at risk. The same tools that make you efficient can also become a backdoor for attackers. ### How the Attack Works The attack exploits a fundamental flaw in how AI agents handle trust. When you ask Claude Code or Codex to scan code, they don't just look at it—they might execute it. The attacker crafts a payload that looks like a benign security patch. The AI, eager to help, approves and runs it. The result? Your machine is now running the attacker's code. It's like hiring a security guard who opens the door for anyone wearing a uniform. ### What You Can Do to Stay Safe - **Limit autonomous mode:** Don't let AI agents approve their own actions. Always require human confirmation before executing code. - **Sandbox your environment:** Run AI agents in isolated containers. Even if they're tricked, the damage stays contained. - **Audit your agents:** Regularly check what your AI tools are doing. Look for unexpected approvals or actions. - **Use trusted sources:** Only scan code from verified repositories. Avoid obscure libraries with few downloads. ### The Bigger Picture This attack highlights a growing problem: we're trusting AI agents with more and more power, but they're still easy to fool. The AI Now Institute's proof-of-concept shows that even top-tier models from Anthropic and OpenAI have this vulnerability. As antidetect browser users, you need to stay ahead of these threats. Your privacy and security depend on it. ### Final Thoughts Don't let this scare you off AI agents entirely. They're powerful tools that can save you hours of work. Just use them wisely. Always double-check what they're doing, and never give them full autonomy. A little caution goes a long way in protecting your digital life.