Why AI Keeps Falling for Prompt Injection Attacks

The report examines why large language models are vulnerable to prompt injection and similar manipulation, arguing that unlike humans—who rely on layered contextual judgment and escalation—LLMs flatten context, are overconfident, and prioritize complying with user requests. It contends that universal safeguards against prompt injection are currently infeasible, risks are amplified with autonomous AI agents, and practical defenses should emphasize narrow scoping, strict escalation, and engineered interruption reflexes while research explores richer world models and context handling.