Scholarlink

Many discussions of AI security risks focus on bad actors—external threats, adversarial inputs, or insider misuse. Some of the most consequential risks stem from something far more mundane: the cognitive limitations of the AI itself. Hallucinations, overconfidence, misalignment, and context drift—these aren't rare edge cases. They are well-documented and, in many cases, expected behaviors of current-generation AI systems.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Cognitive Instability in AI Agents—Security Risks from Misjudgment, Hallucination, and Drift

  • 章节

  • 发表日期：

    2025年

查看全文
  • Rosario Mastrogiacomo

摘要

Many discussions of AI security risks focus on bad actors—external threats, adversarial inputs, or insider misuse. Some of the most consequential risks stem from something far more mundane: the cognitive limitations of the AI itself. Hallucinations, overconfidence, misalignment, and context drift—these aren't rare edge cases. They are well-documented and, in many cases, expected behaviors of current-generation AI systems.