Refusal behavior occurs when a model declines a requested task instead of attempting it. It matters because refusals can reflect safety policies rather than inability, and different models may respond differently to the same request. section 2.4.3
Handle refusals as an operational behavior:
- Test the exact production use case across every target model; documentation describes general policy but may not reveal edge-case triggers. section 2.4.3
- Prepare fallback strategies so an agent can respond safely when one model refuses a request. section 2.4.3
- Define behavioral boundaries in system prompts, including permitted actions, prohibited actions, and limits on legal or financial advice. section 2.3.1
Obfuscated text mistaken for a safety bypass
A GAIA problem presents reversed text that asks for the opposite of “left.” Some models interpret the obfuscation as an attempt to bypass safety filters and refuse instead of answering. section 2.4.3
