Niches · Guardrails & safety filters
Block jailbreaks and prompt injections in what the user types
Two or three companies' docs show a full fit.
What counts as done
A jailbreak or injection in the user turn returns a block or a score above the documented threshold, ordinary requests such as 'Hello' or 'Explain Docker' pass, the detector states which languages it covers, and an unavailable detector is reported rather than treated as a pass.
How agents ask for it: is this prompt a jailbreak
, stop users overriding the system prompt
, add jailbreak detection to our chatbot
Does the job 2
The docs show it meets the test above, and nothing we read says otherwise.
| Listing | Why |
|---|---|
| Amazon Bedrock Guardrails A Amazon Web Services, Inc. | The prompt attack category has its own strength setting, the languages page lists each language as optimised or supported per tier, and failures come back as 429, 500 or 503 rather than a pass. source |
| Google Cloud Model Armor BB Google LLC | Prompt injection and jailbreak filter with a confidence threshold, nine tested languages, inputs under three words return NO_MATCH_FOUND, and a check that could not run returns EXECUTION_SKIPPED or FAILURE rather than a match state. source |
Does the job, with a limit 10
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.
| Listing | Why |
|---|---|
| Azure AI Content Safety (Prompt Shields) B Microsoft Corporation | userPromptAnalysis returns attackDetected true or false with no score or threshold, the page says performance varies outside English, and nothing says how an unavailable check is reported. source |
| Cisco AI Defense Inspection API B Cisco Systems, Inc. | Prompt injection is a named rule with an is_safe verdict, but no score, threshold, language list or behaviour for an unavailable detector was found. source |
| Granite Guardian B International Business Machines Corporation | A jailbreaking criterion judges the last user turn yes or no, but the card says English only, gives no threshold beyond the binary score, and acc-49 found a packaged image without the chat template so the model could not be queried. source |
| Guardrails AI D Guardrails AI, Inc. | A jailbreak validator is listed in the hub catalog, with on_fail actions (exception, fix, noop, custom) to choose what happens, but no languages, threshold defaults or fail behaviour for an unavailable model were found. source |
| Lakera Guard (Check Point AI Guardrails) B Check Point Software Technologies Ltd. | Prompt Defense names 100+ languages and L1 to L4 sensitivity, but the docs do not say what a check that cannot run returns, and the default policy is described as intentionally strict. source |
| LlamaFirewall D Meta Platforms, Inc. | Prompt Guard 2 scores user prompts for jailbreaks, but the docs read name no languages or fail behaviour, the weights need a manual Hugging Face approval that acc-08 was refused, and acc-05 measured 22.8 per cent recall on new injections. source |
| Mistral Moderation API B Mistral AI (RCS Paris 952 418 325) | jailbreaking is a scored category with a settable threshold and block_on_error blocks when moderation fails, but the languages covered are not stated. source |
| NVIDIA NeMo Guardrails BB NVIDIA Corporation | Jailbreak heuristics have documented thresholds but are English only and the length-per-perplexity default catches 31.19 per cent at 7.44 per cent false positives; acc-01 saw the ONNX classifier flag ordinary prompts and acc-02 non-English attacks missed. source |
| OpenAI Guardrails BB OpenAI (as named in the LICENSE copyright line and the PyPI author field) | Jailbreak check returns a confidence against confidence_threshold over up to 10 turns, but no languages are stated and by default a check that fails to run lets the request continue. source |
| Prisma AIRS AI Runtime Security API C Palo Alto Networks, Inc. | Prompt injection detection returns a verdict and a failed or timed-out detector is listed in errors with category error or timeout, but there is no score to threshold and no language list. source |
Doesn't do this job 5
Listings in the category that do something else, with the reason.
| Listing | Why |
|---|---|
| AgentGate D Mohit Singh (individual) | Assesses actions against the session goal; the user's own prompt is not screened for jailbreaks. source |
| Llama Guard 4 D Meta Platforms, Inc. | Detects hazard categories, not jailbreaks or injections, per its own card. source |
| Presidio B Data Privacy Stack (community organisation, no legal entity stated) | No jailbreak or injection detector among its recognisers. source |
| OpenAI Moderation API BB OpenAI OpCo, LLC | No jailbreak or injection category among the harm categories. source |
| ShieldGemma C Google LLC (RDAP registrant of gemma.ai; the Gemma Terms of Use are Google's and the cards name Google as author) | Harm-type policies only; no jailbreak or injection detection. source |
What stops agents finishing
- gated model weights refused at the access request
- classifier artefact flags ordinary greetings as jailbreaks
- English-only detector misses the same attack in another language
Around it
Before: agent-harnesses.always-on-assistant. After: inference.tool-loop.
Where this comes from
Weight 12 of 100 in Guardrails & safety filters: 7 of 50 accounts: NeMo jailbreak-detector false positives and Polish misses, Prompt Guard 2 production use, a vendor-run benchmark, gated-weight refusals and a Lakera adapter that failed on every call
The same as data: /api/v1/use-cases/guardrails.screen-user-prompt.json, or the MCP tool get_use_case with id guardrails.screen-user-prompt. Other jobs in this category: /api/v1/niches?category=guardrails.
For companies
Do agents find, use and choose your tools?
An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.