Cisco AI Defense Inspection API vs Llama Guard 4

Cisco AI Defense Inspection API

Per-connection API keys with expiry, revocation and regeneration, published limits of 60,000 calls a minute, weekly dated release notes and a one-year retention statement suit companies already on Cisco Security Cloud Control. Access needs a subscription bought through sales, with no public price or trial for the API, and the developer changelog has one entry from February 2025.

Llama Guard 4

A single self-hosted model classifies text and multi-image prompts against 14 MLCommons-aligned hazard categories and answers in a few tokens. The weights have not changed since 29 April 2025, download access needs Meta's manual approval, and the licence withholds the grant from individuals and companies based in the European Union.

Cisco AI Defense Inspection API

1. Send the key in X-Cisco-AI-Defense-API-Key to POST /api/v1/inspect/chat on the regional host where the tenant was created. US, Europe and Asia Pacific hosts differ
2. If the connection has a policy, enabled_rules in the request is ignored. Pass enabled_rules only for connections with no policy
3. The API never blocks anything itself. Read is_safe and action in the response and enforce the decision in your own code
4. On 429, wait and retry. The organisation has gone over one million tokens in parallel or 60,000 calls a minute
5. Don't request the Toxicity rule. It was removed on 16 September 2026, and the Safety rules such as Hate Speech, Harassment and Profanity replace it

Llama Guard 4

1. Request access on the Hugging Face page before anything else. Approval is manual, and the form cannot be edited after submission
2. Send only the user turn to check an input, and the user turn plus the model's answer to check an output. The template picks the role from the message count
3. Parse the first line for safe or unsafe and the second for category codes. Set max_new_tokens to about 10 and turn sampling off
4. Do not send an image with no text. Meta says the model is not an image-only classifier, and S14 is skipped when an image is present
5. Pair it with a prompt-attack detector. The card says the model can itself be moved by adversarial or injected text