In this week's episode of This Week in AI Security, Jeremy covers four key developments shaping the AI security landscape.

Jeremy begins by analyzing a GitHub Copilot flaw that exposed an LLM vulnerability similar to the one Jeremy disclosed last week. Researchers were able to use a hidden code comment feature to smuggle malicious prompts into the LLM, allowing them to potentially exfiltrate secrets and source code from private repositories. This highlights a growing risk in how LLMs process different input formats.

Next, we discuss a fascinating research paper demonstrating the effectiveness of data poisoning. The study found that corrupting a model's behavior was possible with as few as 250 malicious documents—even in models with large training sets. By embedding a malicious command that mimicked sudo, researchers could imp ...