Many AI security tools rely on fixed payload libraries.
That's a good starting point.
But attackers rarely repeat the exact same prompt.
They rewrite it.
They obfuscate it.
They change wording.
They manipulate context.
They keep adapting until something works.
Security testing should reflect that reality.
That's why Crucible includes a Mutation Engine that transforms base attack payloads into multiple variations before executing them. Instead of evaluating whether an agent blocks one exact prompt, it evaluates whether the underlying defense remains effective as the attack evolves.
This shifts testing from pattern matching to resilience testing.
Because defending against one prompt is useful.
Defending against an entire family of attacks is far more valuable.
Pytest for AI Agents.
Top comments (0)