Loading content
Attackers exploit the lack of separation between data and instructions to alter an agent's planning, reasoning, or self-evaluation, overriding its intended objective — an extension of prompt injection into long-horizon… In a multi-agent system: An attacker incrementally injects subtly modified sub-goals into a planning agent's context across several turns, d…
Use this profile to understand the building block briefly, place it in the model, and switch to the 360° assessment when needed.
Theoretical construct: explains a term, principle, or mental model.
What organizes, connects, or makes decisions possible.
Intent breaking is the deliberate or accidental alteration of an agent’s goal through manipulated instructions, context, or intermediate results.
The problem grew from prompt-injection and goal-integrity risks in LLM applications; OWASP describes their transfer to agentic planning and execution.
A system keeps the original goal outside the mutable working context. Each planning or tool step is checked against that goal, its limits, and its authorization before taking effect.
Attackers exploit the lack of separation between data and instructions to alter an agent's planning, reasoning, or self-evaluation, overriding its intended objective — an extension of prompt injection into long-horizon…
Primary mitigation: Output Validation / Schema Enforcement at every planning boundary, a Controller for goal-alignment monitoring, and a HITL Approval Gate on plan changes.
For practical classification of Intent Breaking & Goal Manipulation, its concrete application context and the fields of Agentic AI, Risk, Security are decisive.
The concept matters in long conversations, delegation, and autonomous agents. Immutable task metadata, goal checks, action allowlists, and abort paths make gradual goal drift harder.
Where this building block is located in the topic model.
No structure path available.
Explore how this building block connects to concepts, methods, technologies, and tools.
These sources establish the term and its professional meaning.
All direct connections of the current building block in a compact text view.
This classification shows where the building block typically matters, how demanding it is, and what kind of impact it has in the model.
The level within the organization (enterprise, domain, team) at which the AssetBlock is applied.