Why Prompt-Based Defenses Will Fail
You can filter prompts. You can’t filter intent.
Why Prompt-Based Defenses Will Fail
You can filter prompts.
You can’t filter intent.
The dominant strategy in AI security today is simple:
Detect bad inputs. Block bad outputs.
It sounds reasonable.
It’s also fundamentally flawed.
The assumption behind guardrails
Guardrails operate on a core belief:
Harmful intent is visible in text.
So we build:
- classifiers
- filters
- moderation layers
And we expect them to catch misuse.
But intent is not static
Here’s the problem:
Intent is not contained in a single prompt.
It evolves.
A user can:
- start with benign queries
- build context
- gradually shift goals
- disguise harmful intent as legitimate tasks
Each individual message looks safe.
The system as a whole is not.
The multi-step attack problem
Most AI defenses assume:
One prompt → one response → decision
Real systems don’t work like that.
They involve:
- memory
- chaining
- tool usage
- API calls
Attacks happen across steps, not within one input.
Prompt filtering is reactive
Even when guardrails work, they act:
- after input is given
- or after output is generated
By then:
- the model has already processed the request
- the system may have already taken action
The bypass reality
This is why jailbreaks keep working.
Not because models are weak.
But because:
filtering surface-level text cannot capture deeper intent
The real control point
If you want to secure an AI system, you need to control:
- what the user is trying to achieve
- before the system executes anything
Not just:
- what they typed
- or what the model said
A better approach
Instead of asking:
“Is this input allowed?”
We should be asking:
“What is the user trying to do, and should the system do it?”
That requires:
- intent classification
- policy enforcement
- system-level control
The bottom line
Prompt-based defenses will keep improving.
And they will keep failing.
Because they are solving the wrong abstraction.