{
  "listings": [
    {
      "slug": "agentgate",
      "name": "AgentGate",
      "company": "Mohit Singh (individual)",
      "grade": "D",
      "fit": "caveat",
      "why": "history_override takes earlier tool results and the untrusted_influence head scores their effect, but only when a tool call is assessed, with no published accuracy and no input limit.",
      "source": "https://github.com/mohitdebian/AgentGate/blob/main/docs/openapi.yaml",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/agentgate"
    },
    {
      "slug": "amazon-bedrock-guardrails",
      "name": "Amazon Bedrock Guardrails",
      "company": "Amazon Web Services, Inc.",
      "grade": "A",
      "fit": "caveat",
      "why": "Input tagging selects which part of a prompt is evaluated, but there is no field that marks content as untrusted documents, and tool results are not evaluated in tool-use workloads.",
      "source": "https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/amazon-bedrock-guardrails"
    },
    {
      "slug": "azure-ai-content-safety",
      "name": "Azure AI Content Safety (Prompt Shields)",
      "company": "Microsoft Corporation",
      "grade": "B",
      "fit": "caveat",
      "why": "Documents go in their own documents field with per-document results, but the models are primarily trained on English, the limit is 10K characters with no overlap guidance, and acc-26 shows straddling text passing after a cut.",
      "source": "https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/jailbreak-detection",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/azure-ai-content-safety"
    },
    {
      "slug": "cisco-ai-defense-inspection",
      "name": "Cisco AI Defense Inspection API",
      "company": "Cisco Systems, Inc.",
      "grade": "B",
      "fit": "caveat",
      "why": "MCP inspection checks tool responses and resource access as JSON-RPC messages, separate from chat, but no input limit, chunking guidance or indirect-injection accuracy is published.",
      "source": "https://raw.githubusercontent.com/cisco-ai-defense/ai-defense-python-sdk/main/README.md",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/cisco-ai-defense-inspection"
    },
    {
      "slug": "google-model-armor",
      "name": "Google Cloud Model Armor",
      "company": "Google LLC",
      "grade": "BB",
      "fit": "caveat",
      "why": "Text, PDFs and Office files can be scanned up to 65,536 tokens and 4 MB, but there is no field that marks content as untrusted separately from the user prompt.",
      "source": "https://docs.cloud.google.com/model-armor/overview.md.txt",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/google-model-armor"
    },
    {
      "slug": "guardrails-ai",
      "name": "Guardrails AI",
      "company": "Guardrails AI, Inc.",
      "grade": "D",
      "fit": "caveat",
      "why": "The hub lists a prompt injection detector built on Rebuff, but there is no untrusted-content field and the only evidence is the catalog entry.",
      "source": "https://www.guardrailsai.com/hub",
      "confidence": "low",
      "url": "https://www.anchorterminal.com/tools/guardrails-ai"
    },
    {
      "slug": "lakera-guard",
      "name": "Lakera Guard (Check Point AI Guardrails)",
      "company": "Check Point Software Technologies Ltd.",
      "grade": "B",
      "fit": "caveat",
      "why": "tool messages are screened as untrusted content, but no input limit or overlap guidance is given, and acc-23 and acc-22 show framework data headers and a benign web page flagged as attacks.",
      "source": "https://docs.lakera.ai/docs/agent-behavior-defense.md",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/lakera-guard"
    },
    {
      "slug": "llamafirewall",
      "name": "LlamaFirewall",
      "company": "Meta Platforms, Inc.",
      "grade": "D",
      "fit": "caveat",
      "why": "A TOOL role takes tool output separately and Prompt Guard 2 scores it, but acc-06 found the default threshold caught 1 per cent of AgentDojo tool-output attacks, acc-07 that harmless-sounding overrides score near chance, and weights are gated (acc-08).",
      "source": "https://meta-llama.github.io/PurpleLlama/LlamaFirewall/docs/documentation/scanners/prompt-guard-2",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/llamafirewall"
    },
    {
      "slug": "mistral-moderation",
      "name": "Mistral Moderation API",
      "company": "Mistral AI (RCS Paris 952 418 325)",
      "grade": "B",
      "fit": "caveat",
      "why": "The jailbreaking category can score any text, but there is no field marking content as untrusted, no input limit is stated and indirect attacks are not described.",
      "source": "https://docs.mistral.ai/studio/safety-moderation.md",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/mistral-moderation"
    },
    {
      "slug": "nemo-guardrails",
      "name": "NVIDIA NeMo Guardrails",
      "company": "NVIDIA Corporation",
      "grade": "BB",
      "fit": "caveat",
      "why": "PII and third-party rails can run on retrieved content, but the built-in injection detection is YARA rules for code, SQL and template injection in output, not prompt injection in tool results.",
      "source": "https://docs.nvidia.com/nemo/guardrails/latest/configure-guardrails/guardrail-catalog/agentic-security.md",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/nemo-guardrails"
    },
    {
      "slug": "openai-guardrails",
      "name": "OpenAI Guardrails",
      "company": "OpenAI (as named in the LICENSE copyright line and the PyPI author field)",
      "grade": "BB",
      "fit": "caveat",
      "why": "Prompt injection detection checks function call outputs against the user's goal and is benchmarked on AgentDojo traces, but its table shows recall 0.000 at 1 per cent false positives for every model, below the measured bar.",
      "source": "https://openai.github.io/openai-guardrails-python/ref/checks/prompt_injection_detection/",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/openai-guardrails"
    },
    {
      "slug": "prisma-airs",
      "name": "Prisma AIRS AI Runtime Security API",
      "company": "Palo Alto Networks, Inc.",
      "grade": "C",
      "fit": "caveat",
      "why": "tool_event carries tool input and output separately from the prompt and prompt injection is a detection service, but no input limit, chunking guidance or benign rate on tool output is published.",
      "source": "https://raw.githubusercontent.com/PaloAltoNetworks/pan.dev/master/openapi-specs/prisma-airs/scan/scan-service_latest.yaml",
      "confidence": "medium",
      "url": "https://www.anchorterminal.com/tools/prisma-airs"
    },
    {
      "slug": "granite-guardian",
      "name": "Granite Guardian",
      "company": "International Business Machines Corporation",
      "grade": "B",
      "fit": "no",
      "why": "Documents can be passed separately, but the pre-baked criteria judge the last user or assistant turn and none targets instructions hidden in documents or tool output.",
      "source": "https://huggingface.co/ibm-granite/granite-guardian-4.1-8b",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/granite-guardian"
    },
    {
      "slug": "llama-guard",
      "name": "Llama Guard 4",
      "company": "Meta Platforms, Inc.",
      "grade": "D",
      "fit": "no",
      "why": "No prompt-injection category; the card says the model can itself be moved by injected text and points to Prompt Guard 2.",
      "source": "https://raw.githubusercontent.com/meta-llama/PurpleLlama/main/Llama-Guard4/12B/MODEL_CARD.md",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/llama-guard"
    },
    {
      "slug": "microsoft-presidio",
      "name": "Presidio",
      "company": "Data Privacy Stack (community organisation, no legal entity stated)",
      "grade": "B",
      "fit": "no",
      "why": "No prompt-injection detection; it finds PII entities in text it is given.",
      "source": "https://presidio.dataprivacystack.org/supported_entities/",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/microsoft-presidio"
    },
    {
      "slug": "openai-moderation",
      "name": "OpenAI Moderation API",
      "company": "OpenAI OpCo, LLC",
      "grade": "BB",
      "fit": "no",
      "why": "No prompt-injection category; a clean result says nothing about instructions hidden in tool output.",
      "source": "https://developers.openai.com/api/docs/guides/moderation",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/openai-moderation"
    },
    {
      "slug": "shieldgemma",
      "name": "ShieldGemma",
      "company": "Google LLC (RDAP registrant of gemma.ai; the Gemma Terms of Use are Google's and the cards name Google as author)",
      "grade": "C",
      "fit": "no",
      "why": "No prompt-injection or jailbreak policy in either generation.",
      "source": "https://gemma.ai/models/shieldgemma/model-card-v1.md",
      "confidence": "high",
      "url": "https://www.anchorterminal.com/tools/shieldgemma"
    }
  ],
  "marks": "yes: the docs show it does the job as the success test says. caveat: it does the job with a limit an agent should know. no: it doesn't do this job. unchecked: we couldn't read it.",
  "meta": {
    "attribution": "Anchor Terminal (https://www.anchorterminal.com)",
    "attributionRequired": true,
    "credit": "CC BY 4.0, credit Anchor Terminal and link the page or API address you took it from (https://www.anchorterminal.com/terms/#data)",
    "docs": "https://www.anchorterminal.com/docs/",
    "generatedAt": "2026-11-16",
    "license": "CC-BY-4.0",
    "method": "https://www.anchorterminal.com/benchmark/",
    "methodology": "0.8",
    "openapi": "https://www.anchorterminal.com/openapi.json",
    "preview": false,
    "run": "2026-10-01",
    "runLabel": "October 2026 research run",
    "terms": "https://www.anchorterminal.com/terms/#data"
  },
  "useCase": {
    "accounts": 9,
    "after": [
      "inference.tool-loop",
      "human-in-the-loop.approve-action"
    ],
    "area": "models",
    "asks": [
      "is there a prompt injection in this page",
      "check these search results before the agent uses them",
      "screen uploaded files for hidden instructions",
      "scan the email body before the agent acts on it"
    ],
    "before": [
      "web-scrapers.page-to-markdown",
      "search.fetch-url",
      "mailbox-access.find-and-read"
    ],
    "blockers": [
      "default threshold catches almost none of the indirect attacks",
      "agent framework's own data-wrapper text flagged as an attack",
      "document attack detection off by default"
    ],
    "capabilities": [
      "guard.injection",
      "guard.indirect-injection"
    ],
    "category": "guardrails",
    "categoryName": "Guardrails \u0026 safety filters",
    "checked": "2026-11-15",
    "companies": {
      "fit": 0,
      "fitWithLimit": 12,
      "tagged": 16
    },
    "competition": "none",
    "fits": null,
    "fitsWithLimit": [
      "agentgate",
      "amazon-bedrock-guardrails",
      "azure-ai-content-safety",
      "cisco-ai-defense-inspection",
      "google-model-armor",
      "guardrails-ai",
      "lakera-guard",
      "llamafirewall",
      "mistral-moderation",
      "nemo-guardrails",
      "openai-guardrails",
      "prisma-airs"
    ],
    "id": "guardrails.screen-untrusted-content",
    "listings": {
      "caveat": 12,
      "no": 5,
      "notRead": 0,
      "tagged": 16,
      "unchecked": 0,
      "yes": 0
    },
    "name": "Check tool results, pages and uploads for hidden instructions before the agent reads them",
    "status": "draft",
    "success": "The content is submitted in a field marked as untrusted, separate from the user prompt, attacks phrased as harmless requests are scored as well as 'ignore previous instructions', the agent's own framing text is not flagged, and text longer than the input limit is split with overlap. Measured: at least 90 per cent of AgentDojo tool-output injections caught at no more than 2 per cent false alarms on benign tool output.",
    "url": "https://www.anchorterminal.com/api/v1/use-cases/guardrails.screen-untrusted-content.json",
    "weight": 16,
    "weightBasis": "9 of 50 accounts: Prompt Guard 2 benchmarks on AgentDojo and documents, Prompt Shields wired into upload and tool-result paths in three repos, and Lakera false positives on a scraped page and a subagent result; the buyer checks put this first, but the weight follows the accounts"
  }
}
