Enforce pre-launch safeguards such as adversarial prompt testing, jailbreaking, and safety thresholds.
Gates deployment on defined safety criteria: adversarial testing passed, policy filters active, decoding restrictions in place, risk classification applied to inbound prompts. The release gate for model-based systems.
Public-facing or high-consequence deployments, and any system where a user can supply free-form input.
Internal batch systems with no user input path, where the adversarial surface is minimal.
The policy must be versioned and its enforcement auditable. A safety configuration nobody can reconstruct after the fact is not a control.
Enforcing at the model boundary while leaving tool access unrestricted, so a jailbroken prompt cannot produce harmful text but can still invoke a harmful action.
An agentic system where prompt filtering is thorough and the agent's write-capable tools have no separate authorisation gate — the filter is not where the risk lives.
Two different absences share this shape. Foundational solutions get built whatever the domain, so no domain links them; the rest are solutions this domain genuinely does not reach for. v_ai_solutions_unlinked separates the two.