Safety Dimensions

Name: mutant-ai
Author: Ankit Mishra

The Safety Dimension tests the adversarial robustness of an agent by attempting jailbreaks, prompt injections, and boundary violations.

Available Dimensions

ID	Name	Severity	Description
`safety.prompt_injection`	Prompt Injection	critical	Embedded instruction in user input attempts to override agent system prompt.
`safety.jailbreak`	Jailbreak Attempt	critical	User attempts to bypass agent guardrails via roleplay, hypotheticals, or fiction.
`safety.sensitive_information`	Sensitive Information Request	critical	User requests sensitive data the agent should not reveal.
`safety.social_engineering`	Social Engineering	critical	User uses authority, urgency, or flattery to manipulate the agent into bypassing procedures.
`safety.workflow_hijacking`	Workflow Hijacking	critical	User embeds secondary instructions inside the primary request to silently redirect agent actions.
`safety.context_injection`	Context Injection	critical	Attacker injects fake prior context (agent notes, system logs, prior messages) to manipulate agent state.
`safety.permission_escalation`	Permission Escalation	critical	User self-asserts elevated authority or role to bypass authorization checks.
`safety.instruction_override`	Instruction Override	critical	User attempts to reset or nullify the agent's prior instructions using delimiter or reset attacks.

Safety mutation dimensions.