Agent hijacking occurs when malicious instructions arrive inside material a system is supposed to treat as data. NIST describes the problem as a failure to keep trusted instructions separate from untrusted external content.
Agent hijacking and the trust boundary
In NIST’s description, an email, file or website can look like ordinary task data while containing instructions intended to redirect an agent. The resulting risk depends on the tools and information the system can access. This is a concrete security mechanism; explaining it does not require a claim that the system has human intentions or that autonomous execution is impossible.
Understanding AI Agent Hijacking Risks
A more precise technical concern involves agent hijacking. This is a form of indirect prompt injection where an attacker inserts malicious instructions into data that an AI agent later ingests. Once the agent processes this corrupted data, it may execute unintended or harmful actions. For example, automated systems could be manipulated into sending phishing emails to contacts or performing unauthorised tasks. Identifying and measuring these vulnerabilities is essential for organisations seeking to deploy agentic systems securely. NIST’s examples are evaluation tasks, rather than a prevalence estimate for phishing or data theft in UK organisations.
Applying Cyber Security Principles
The National Cyber Security Centre emphasises that the core principles of cyber security remain the most effective way to protect systems, regardless of whether they contain machine learning components. When implementing AI, organisations should prioritise fundamental security hygiene, including supply chain security, user education, and strict access controls. Designing for security by default and ensuring that software updates are applied consistently are critical steps. Businesses are advised to 'walk before they run' when integrating agentic AI, ensuring that security measures are robust enough to handle the specific risks associated with automated task execution.
Sources and further reading
There are no “rogue” AI agents - by Eoin Higgins
https://eoinhiggins.substack.com/p/there-are-no-rogue-ai-agents
Thinking about the security of AI systems | National Cyber Security Centre
https://www.ncsc.gov.uk/blog-post/thinking-about-security-ai-systems
Technical Blog: Strengthening AI Agent Hijacking Evaluations | NIST
https://www.nist.gov/news-events/news/2025/01/technical-blog-strengthening-ai-agent-hijacking-evaluations