Jailbreaks and benign controls
Separate unsafe compliance from unnecessary refusal. Define attacker access, score response behavior and test legitimate requests alongside attacks.
Read the Cortega white paper →Cortega Red Team
Run adaptive conversations, poisoned-memory tests and unauthorized tool-access attempts. Review what the attacker controlled, how the agent responded and the evidence behind each finding.
Explore the test methods →The basis of the tests
Test direct jailbreaks, adaptive conversations, long-context pressure and instructions embedded in tool results. The Cortega white-paper library explains each method, its research basis and how to evaluate it.
Separate unsafe compliance from unnecessary refusal. Define attacker access, score response behavior and test legitimate requests alongside attacks.
Read the Cortega white paper →How Crescendo, GOAT, PAIR and TAP inform tests that revise their approach after the target responds.
Read the Cortega white paper →Test whether repeated examples and their position in context change the agent’s behavior, with matched benign workloads.
Read the Cortega white paper →Evaluate whether untrusted tool results or stored context can redirect an agent into an unauthorized action.
Read the Cortega white paper →Published model-safety evidence
The application can inspect 640 saved samples from the historical two-arm model evaluation. This imported benchmark is separate from application campaign results.
