AI Security Topics

Direct and indirect prompt injection against LLM applications and AI agents: instructions hidden in user input, documents, web pages, emails or tool output that change what the model does.

Security of the Model Context Protocol (MCP): vulnerable or malicious MCP servers, tool poisoning, over-broad tool permissions and exposed MCP endpoints.

Vulnerabilities and compromises in AI frameworks, model hubs, agent tooling, packages and training data, plus model and data poisoning.

AI agents with more permissions, tools or autonomy than their task needs, and incidents where agents took harmful or unapproved actions.

Attackers using AI: LLM-written malware, AI-generated phishing and voice-cloning scams, and automated vulnerability discovery and exploitation.