New Prompt Injection Techniques Cover Agent-Planning Attacks
CrowdStrike, Thursday, October 29th, 2026
CrowdStrike's updated Prompt Injection Taxonomy adds agent-planning attacks, bringing it to 222 documented techniques.
CrowdStrike says prompt injection has moved beyond telling a chatbot to ignore its rules to manipulating how agents plan and how models decide which rules apply.
The new version of its public Prompt Injection Taxonomy adds techniques aimed at autonomous agents, for a total of 222.
Attackers can shape the context an agent plans from so a restricted action looks like the logical next step, and step-by-step checks may miss it.
An example shows a tampered test utility steering a coding agent into sending local config files to an attacker-run endpoint.