The AI ​​security nightmare is here and looks suspiciously like a lobster

A hacker tricked a popular AI-powered encryption tool into installing OpenClaw — the Open source viral AI agent OpenClaw That “actually does things” – absolutely everywhere. Funny as a trick, but it’s a sign of what’s to come as more and more people let standalone software use their computers for them.

The hacker exploited a vulnerability in Cline, an open source AI encryption proxy popular among developers, that security researcher Adnan Khan had exploited. appeared Just days ago as a proof of concept. Simply put, Klein’s workflow used Anthropic’s Claude technology, which could be fed deceptive instructions and made to do things they shouldn’t do, a technique known as instantaneous injection.

The hacker used his access to sneak past instructions to automatically install software on users’ computers. They could have installed anything, but they chose OpenClaw. Fortunately, the clients were not activated upon installation, otherwise this would have been a completely different story.

It’s a sign of how quickly things can unravel when AI agents are given control of our computers. It may seem like a clever play on words — one set Grooming chatbots to commit crimes using poetry – But in a world Increasingly independent software,Instantaneous injections present enormous security risks that are ,very difficult to defend against. Recognizing this, some companies are instead shutting down what AI tools can do if they are compromised. OpenAI, for example, recently introduced a new application Lock mode For ChatGPT it prevents it from giving out your data.

Obviously, it’s harder to protect against rapid injections if you ignore researchers who secretly report flaws to you. Khan said he warned Klein about the weaknesses weeks before publishing his findings. The exploit was only fixed after he called them out publicly.