Cybersecurity

Data poisoning is the deliberate injection of adversarial content into a model’s training data or a tool’s metadata so the LLM learns or obeys malicious instructions. In agentic systems that load third-party tools from MCP (model-connected platform) servers, poisoned tool descriptions or docstrings can trick an LLM into leaking secrets, executing harmful actions, or behaving as a covert proxy for attackers.

There is growing threat of AI enhanced supply chain attacks targeting Python packages, FOSS software and why sometimes writing your own isolated, network-free software might be our best defense