PocketClawvol. 1 · 2026

Prompt injection

Attack technique where adversarial input embedded in a document or webpage hijacks an agent's behaviour.

Prompt injection is the dominant agentic attack vector in 2026. A malicious document, webpage, or tool description can attempt to override the agent's system prompt or redirect its goals. Defences include input sanitisation, capability scoping, never letting the agent act autonomously on adversarial input, and human-in-the-loop approvals.

Related terms

AI agentSandboxTool call

Found a definition that's wrong, dated or could be sharper? Email us — we update with attribution unless you'd rather we didn't.