# Enforce one guardrail policy on every request through a gateway

> Enforce one guardrail policy on every request through a gateway: 0 companies fully do it, 9 with a limit, 25 carry the tag. Every listing in Guardrails & safety filters with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/guardrails.gateway-enforcement/
- Markdown: https://www.anchorterminal.com/niches/guardrails.gateway-enforcement/index.md (~1,800 tokens)
- Slim: https://www.anchorterminal.com/niches/guardrails.gateway-enforcement/index.min.md (~380 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/guardrails.gateway-enforcement/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

No company's docs show a full fit yet. The listings below do the job with a limit, or not at all.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 0 | 9 | 25 | 19 | 11 |

## What counts as done

Every request route the gateway serves (chat, Responses, Messages, batch, streaming) is screened, the whole conversation is checked rather than only the last message, a check that did not run is logged as not run, and a block returns a reason the caller can act on.

Measured, for our own tests: the check adds under 150 ms at the median for a 2,000-token request.

How agents ask for it: put a guardrail in front of all our model traffic; make every team's calls go through the same checks; add Bedrock or Model Armor screening to our LiteLLM proxy

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Amazon Bedrock Guardrails (Amazon Web Services, Inc.)**: ApplyGuardrail screens text and image content blocks with per-policy assessments and errors as HTTP codes, but tool-use fields are skipped (acc-18) and a gateway integration dropped images from the payload (acc-19).
- **Azure AI Content Safety (Prompt Shields) (Microsoft Corporation)**: shieldPrompt takes one userPrompt and up to five documents, not a message history, within 10K characters; no fail-mode or not-run status is documented, and document attack detection is off by default in Foundry.
- **Cisco AI Defense Inspection API (Cisco Systems, Inc.)**: Chat inspection takes the whole messages array, HTTP inspection covers raw requests, and the SDK has a gateway mode with enforce or monitor; but the API never blocks itself and no fail-mode or not-run status was found in what was read.
- **Google Cloud Model Armor (Google LLC)**: Each prompt or response is inspected as a single-turn request with no conversation history; a skipped or failed check is reported (EXECUTION_SKIPPED, invocation_result FAILURE), but acc-20 and acc-21 show Responses traffic unscreened and batches failing in a gateway.
- **Lakera Guard (Check Point AI Guardrails) (Check Point Software Technologies Ltd.)**: Only the last interaction is screened, earlier turns are context; content type text only, so images are not checked. acc-22 saw a 403 with no detector named and acc-24 a v1 body sent to v2 that blocked all traffic.
- **Mistral Moderation API (Mistral AI (RCS Paris 952 418 325))**: Inline guardrails with per-category thresholds, action block and block_on_error cover Mistral's chat, conversations and agents only; the conversational endpoint classifies the last turn, and text only.
- **NVIDIA NeMo Guardrails (NVIDIA Corporation)**: The server fronts any model with an OpenAI-compatible endpoint and /v1/checks, and tool rails fail closed, but acc-03 found only the last user message checked and acc-04 that a cut-off verdict passes as safe.
- **OpenAI Guardrails (OpenAI (as named in the LICENSE copyright line and the PyPI author field))**: A client wrapper rather than a gateway; checks fail open unless raise_guardrail_errors=True, and the prompt injection check's own table gives a 1,481 ms median for the default model, far above the 150 ms bar.
- **Prisma AIRS AI Runtime Security API (Palo Alto Networks, Inc.)**: Every response carries action allow or block, a category, and errors for detection services that failed or timed out; but only the last element of contents is scanned, earlier turns are context, and content is text only.

## Couldn't check

The docs were closed to us, or silent. These carry no mark either way.

- **Guardrails AI (Guardrails AI, Inc.)**: The docs site returned 404, and the repository pages read do not say which request routes the API server screens, whether history is checked, or how a validator that fails to run is reported.

- not this job: AgentGate (Mohit Singh (individual)): Sits between an agent and its tools, assessing tool invocations; it does not screen chat, Responses, Messages or batch model traffic.
- not this job: Granite Guardian (International Business Machines Corporation): An 8B judge model the owner runs, one criterion per call; not a gateway or policy layer.
- not this job: Llama Guard 4 (Meta Platforms, Inc.): A classifier model the owner hosts, not a policy layer; no injection or PII checks, and the card sends readers to Prompt Guard 2 for prompt attacks.
- not this job: LlamaFirewall (Meta Platforms, Inc.): An in-process Python library that scans messages the agent hands it; it is not a proxy and has no route coverage, moderation or PII masking.
- not this job: Presidio (Data Privacy Stack (community organisation, no legal entity stated)): A PII detection and anonymisation library with no gateway, injection or moderation checks of its own; route coverage belongs to the proxy that calls it, and acc-15, acc-16 and acc-43 show LiteLLM's Presidio path missing tool arguments and streams.
- not this job: OpenAI Moderation API (OpenAI OpCo, LLC): A harm classifier, not a policy layer: no injection, PII or conversation-wide checks, and inline moderation on a streamed generation arrives only after the full output.
- not this job: ShieldGemma (Google LLC (RDAP registrant of gemma.ai; the Gemma Terms of Use are Google's and the cards name Google as author)): An open-weight classifier the owner runs, one policy per call; no injection or PII checks and no gateway.

## What stops agents finishing

- guardrail reads only chat-completions messages, so Responses or batch traffic passes unscreened
- API version change makes every check fail and the gateway blocks all traffic
- image parts dropped from the guardrail payload while still sent to the model

## Around it

Before: inference.failover-routing. After: agent-observability.gateway-logging, human-in-the-loop.audit-decisions.

## Where this comes from

Weight 19 of 100 in Guardrails & safety filters: 11 of 50 accounts: LiteLLM, Otari, Future AGI and OpenClaw-plugin issues where guardrails skipped a route, dropped images, failed open or blocked everything, plus a NeMo gateway integrator's latency benchmark

As data: https://www.anchorterminal.com/api/v1/use-cases/guardrails.gateway-enforcement.json, or the MCP tool get_use_case with id guardrails.gateway-enforcement. Draft, read 2026-11-15 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
