In Halo you can
The agent can try whatever it wants. The runtime just says no.
halo
policy Safe : Block(PromptInjection) & Redact(APIKey) policy Budget : MaxSpend(0.25) agent Writer [anthropic/claude-haiku-4-5] requires Safe, Budget
Hot take: if your ai/agent safety lives only in the prompt, you don’t have safety.
commit c7bed25