Promptjacking

Prompt injection is the technical term security researchers use. Promptjacking is the plain-English term entering general builder and media vocabulary to describe the same family of attacks: someone crafts text inputs that cause an AI system to abandon its intended behavior and follow instructions it was not supposed to follow. The 'jacking' framing is deliberate: like hijacking a vehicle mid-journey, the system keeps running but the destination changes.

The term covers a spectrum of techniques. Direct promptjacking happens when a user types instructions designed to override the system prompt, such as 'ignore all previous instructions.' Indirect promptjacking happens when malicious instructions are hidden inside content the AI retrieves or reads: a web page, a PDF attachment, a tool response. In agentic systems where the AI has access to email, files, calendars, and APIs, indirect promptjacking is particularly dangerous because the attack surface extends to any content the agent can touch.

Promptjacking entered mainstream tech vocabulary in 2026 as AI agents became common in enterprise settings and the consequences of a successful hijack grew beyond embarrassing outputs to actual data exfiltration, unauthorized actions, and financial transactions. Builders should treat promptjacking as a first-class threat model when designing any agent that reads external content, integrates with third-party services, or operates with elevated permissions. Mitigations include input validation, scoped permissions, output monitoring, and running red-team evals that simulate real hijack scenarios.