LLM firewalls and AI guardrails
An LLM firewall or guardrail layer sits between your users, your agents and your models. It reads each prompt and response and blocks prompt injection, jailbreaks and data leaks before they land. 38 of 56 vendors document it. 9 of them let you run it in your own environment, and NVIDIA NeMo Guardrails is open source.
Latency is the number buyers ask for first, and few publish it: Check Point AI Security claims under 50 ms, Fiddler AI claims under 80 ms and NeuralTrust claims under 100 ms. All three are vendor claims, not benchmarks.
38 vendors that document runtime guardrails
Checked 11 Nov 2026| Vendor | What it documents | Also covers |
|---|---|---|
| A10 Networks TrojAI | TrojAI Defend blocks prompt injection, jailbreaks, data leakage, and toxic content. 1 | Agent security, Red teaming |
| Akamai Firewall for AI | Detects and blocks prompt injections, jailbreaks and harmful queries before they reach the model; inspects inputs and outputs in real time. 1 | Data protection |
| Alice (ActiveFence) | WonderFence inspects inputs and outputs and blocks or redacts policy violations, including injection and jailbreak attempts. 1 | Red teaming, Data protection |
| Amazon Bedrock Guardrails | Configurable content/word filters, prompt attack detection, denied topics, and hallucination checks. 1 | Data protection |
| Arthur | Provides real-time guardrails and security-policy enforcement. 1 | Agent security, AI-SPM |
| Aurascape | Enforces continuous runtime guardrails. 1 | Agent security, Red teaming, AI-SPM, Data protection |
| Cato Networks AI Security | Aim AI Firewall enforces security policies on runtime interactions with private AI apps and agents. 1 | AI-SPM |
| Check Point AI Security | Detects prompt injection, jailbreaks, and data exfiltration. 1 | Agent security, Red teaming |
| Cisco AI Defense | Network-embedded guardrails block attacks and harmful responses in production. 1 | Red teaming, AI-SPM, Model scanning, Data protection |
| Cloudflare AI Security for Apps | Detects prompt injection and unsafe or custom topics in incoming JSON prompts; WAF rules apply blocking or challenges. 1 | Data protection |
| Coralogix AI Guardrails | Detects, blocks, rewrites, or flags unsafe prompts and responses. 1 | Agent security, Data protection |
| DeepKeep | AI Firewall supplies runtime detection and guardrails. 1 | Agent security, Red teaming, Model scanning |
| Enkrypt AI | Real-time guardrails stop threats and adapt using prompts. 1 | Agent security, Red teaming |
| F5 AI Guardrails | Defends against prompt injection, jailbreaks, data leakage, and harmful outputs. 1 | Agent security, Red teaming |
| Fiddler AI | Enforces policy inline on the agent request and response path. 1 | AI-SPM, Data protection |
| Fortinet FortiAIGate | Inspects text, images, audio, video, and code for prompt injection, jailbreaks, and data leaks. 1 | Agent security, Data protection |
| Google Cloud Model Armor | Blocks prompt injection and jailbreak attempts. 1 | Data protection |
| Gray Swan | Cygnal classifies and blocks adversarial inputs and unsafe agent outputs in production. 1 | Agent security, Red teaming |
| HiddenLayer | Uses policy controls against misuse, data leakage, and adversarial attacks. 1 | Agent security, Red teaming, Model scanning |
| Lasso Security | Enforces policies inline through proxy, API, or AI Gateway. 1 | Red teaming, AI-SPM |
| Microsoft AI security | Prompt Shields scans text for user-input attacks on LLMs. 1 | Agent security, AI-SPM, Data protection |
| Netskope | Inline module protects against prompt injection, jailbreaks, and unsafe content. 1 | Agent security, Red teaming |
| NeuralTrust | Centrally enforces prompt/content filters, DLP, and tool-use permissions. 1 | AI-SPM |
| Noma Security | Can stop destructive actions, steer agents, alert humans, mask data, or trigger a kill switch. 1 | Agent security, Red teaming, Model scanning |
| NVIDIA NeMo Guardrails | Input, retrieval, dialog, execution, and output rails support policy enforcement. 1 | Agent security |
| Operant AI | Semantic Firewall makes inline allow, block, or redact decisions using agent-action intent. 1 | Agent security, AI-SPM |
| Palo Alto Networks Prisma AIRS | Enforces runtime safeguards against manipulation, data exposure, and unsafe actions. 1 | Agent security, Red teaming, AI-SPM, Model scanning |
| Pillar Security | Provides runtime guardrails and agent runtime protection. 1 | Red teaming, AI-SPM |
| Proofpoint AI Security | Enforces context-aware policies during AI interactions. 1 | Agent security, AI-SPM, Data protection |
| Qualys TotalAI | Provides live guardrails and a prompt firewall with container behavior protection. 1 | Red teaming, AI-SPM |
| SentinelOne Prompt Security | Blocks adversarial prompts and scrubs sensitive outputs through a runtime AI firewall. 1 | Agent security, Red teaming, AI-SPM |
| Straiker | Defend AI blocks identity abuse, memory poisoning, data exfiltration, and resource exploitation. 1 | Agent security, Red teaming, AI-SPM |
| Tenable One AI Exposure | Detects and stops prompt injection and jailbreak attempts. 1 | AI-SPM |
| Trend Micro TrendAI Vision One | AI Guard inspects prompts and responses for harmful outputs, injection, and sensitive-data leakage. 1 | Red teaming, AI-SPM |
| WitnessAI | Blocks jailbreaks and prompt injection at runtime. 1 | Agent security, Red teaming, AI-SPM |
| Wiz | Detects prompt injection, rogue agents, and malicious AI behavior at runtime. 1 | AI-SPM, Data protection |
| Zenity | AI detection and response blocks unsafe agent actions in real time or shuts the agent down. 1 | Agent security, AI-SPM |
| Zscaler AI Security | Runtime guardrails and prompt hardening protect data and block malicious interactions. 1 | Red teaming, AI-SPM |
What to ask before you shortlist
- What does it inspect? Text alone is table stakes. Fortinet FortiAIGate documents text, images, audio, video and code; Google Model Armor documents PDFs and other unstructured documents.
- Where does it run? Self-host or private cloud: A10 Networks TrojAI, F5 AI Guardrails, Fiddler AI, Fortinet FortiAIGate, Netskope, NeuralTrust, NVIDIA NeMo Guardrails, Pillar Security and Trend Micro TrendAI Vision One.
- What latency do you add? Ask for p95 on your own prompt sizes, with and without streaming.
- Which frameworks does it map to? 12 of these vendors map controls to OWASP, MITRE ATLAS, NIST, GDPR or the EU AI Act. The mapping is on each vendor page, with its source.
- Does it cover agents? A guardrail on text does not stop a bad tool call. Check the agent security list.