Seonglae ChoAgentic Misalignment: How LLMs could be insider threats
The experiment tested whether AI agents would choose harmful actions when faced with replacement threats or goal conflicts. Most models engaged in blackmail and corporate espionage at significant rates under threat or goal conflict conditions, despite being aware of ethical constraints. This suggests that simple System Prompt guidelines are insufficient for prevention, highlighting the need for human oversight, Mechanistic interpretability Activation Steering, real-time monitoring(AI Observability), and transparency before granting high-risk permissions.
Agentic Misalignment: How LLMs could be insider threats
New research on simulated blackmail, industrial espionage, and other misaligned behaviors in LLMs
https://www.anthropic.com/research/agentic-misalignment
Task Eagerness Inner Alignment Goal Misgeneralization
The core issue is exceeding granted permissions, taking irreversible actions, and acting without being asked. OWASP's Excessive Agency captures the same concept: overaction.