Security & ITBest AI Runtime Security and Monitoring Tools
Also known as: AI runtime monitoring tools · AI runtime protection tools · AI SPM tools · AI security posture management tools · AI agent security tools
AI runtime security tools watch AI apps, models and agents while they run in production, and stop or flag attacks as they happen: prompt injection, data leaking out through a chatbot, an agent misusing a tool, or a rogue agent nobody approved. Most also include AI security posture management (AI-SPM), which keeps an inventory of every model, agent, MCP server and AI service an organisation uses and flags risky set-ups before anyone attacks them.
These are products for security teams, not developers. Our AI guardrails ranking covers the filters developers add to a single app, and our LLM observability ranking covers tracing, quality and cost. Here we ranked 12 platforms that give a security team detection, response, posture and governance across all of an organisation's AI, scored on six criteria as of 9 October 2026. Almost every vendor has changed owner or added agent features in the past year, and no independent lab has yet published comparative test results, so we flag which figures are vendor claims.
Quick answerPalo Alto Networks Prisma AIRS is the best AI runtime security platform for most large organisations in 2026. It combines runtime protection for apps and agents, AI-SPM across AWS, Azure and Google Cloud, model scanning, red teaming and an AI Agent Gateway. Pick Wiz if posture across several clouds matters most, Zenity or Noma Security to secure AI agents in SaaS tools and on laptops, and Microsoft Defender for Cloud if you build on Azure and want published per-token prices.
Top picks at a glance
Best overallPalo Alto Networks Prisma AIRS
Runtime protection, AI-SPM, model scanning, red teaming and an agent gateway in one platform.
Best for AI-SPMWiz AI-SPM and AI-APP
Agentless AI inventory and attack paths across AWS, Azure, Google Cloud and Oracle Cloud.
Best for AI agent securityZenity
Widest coverage of SaaS, cloud and coding agents, with in-line blocking and public usage prices.
Best for agents on employee devicesNoma Security
Registry and runtime control for Claude Code, Cursor, Codex and MCP servers on laptops.
Best for AzureMicrosoft Defender for Cloud (AI threat protection and AI-SPM)
Native posture and threat alerts for Azure AI at $0.0008 per 1,000 tokens scanned.
Scoreboard
Scores are out of 10. The overall score is the weighted average of the criteria below.
| # | Tool | Overall | Runtime detection & response | AI-SPM & discovery | Agent & MCP coverage | Deployment & integrations | Evidence & transparency | Pricing transparency | Price from | Best for |
|---|
| 1 | Palo Alto Networks Prisma AIRS Palo Alto Networks | 8.5 | 9.2 | 9.0 | 9.3 | 8.5 | 8.0 | 5.0 | Custom | Large organisations that want runtime protection, AI-SPM and agent control in one platform |
| 2 | Wiz AI-SPM and AI-APP Wiz (Google Cloud) | 8.3 | 7.8 | 9.6 | 8.3 | 8.8 | 8.0 | 6.5 | $24,000/year (100 workloads) | Cloud security teams that need AI inventory and attack paths across several clouds |
| 3 | Zenity Zenity | 8.3 | 8.5 | 8.0 | 9.5 | 8.0 | 7.5 | 7.5 | Usage-based ($16 per million tokens) | Securing AI agents built in Microsoft, Salesforce, ServiceNow and coding tools |
| 4 | Noma Security Noma Security | 8.2 | 8.3 | 8.5 | 9.0 | 8.0 | 7.3 | 7.0 | Usage-based ($8 per million tokens) | Governing coding agents, MCP servers and skills on employee laptops as well as cloud agents |
| 5 | Check Point AI Defense Plane (Lakera) Check Point Software Technologies | 8.0 | 8.8 | 7.3 | 8.0 | 8.3 | 7.5 | 7.5 | $74,000/year (10M requests) | Fast in-line blocking for AI apps and agents, with a self-hosted option |
| 6 | Cisco AI Defense Cisco | 7.9 | 8.5 | 8.0 | 8.5 | 8.3 | 8.0 | 4.5 | Custom | Teams that want an agent SDK developers will adopt, backed by a network security vendor |
| 7 | Microsoft Defender for Cloud (AI threat protection and AI-SPM) Microsoft | 7.8 | 7.5 | 8.5 | 6.5 | 7.5 | 8.5 | 9.5 | Usage-based ($0.0008 per 1,000 tokens) | Teams building on Azure that want native alerts and published per-token prices |
| 8 | CrowdStrike Falcon AIDR (Falcon Guardian) CrowdStrike | 7.5 | 8.8 | 6.8 | 8.5 | 8.0 | 7.0 | 3.5 | Custom | Falcon customers who need to find and control AI agents on laptops and desktops |
| 9 | SentinelOne Prompt Security SentinelOne | 7.5 | 8.5 | 7.8 | 8.3 | 8.0 | 6.5 | 3.5 | Custom | Stopping data leaks through employee AI use and code assistants |
| 10 | HiddenLayer HiddenLayer | 7.4 | 8.3 | 7.5 | 8.0 | 7.5 | 7.8 | 3.0 | Custom | Regulated enterprises protecting models, agents and AI coding agents with an independent vendor |
| 11 | Pillar Security Pillar Security | 7.4 | 8.0 | 8.0 | 8.3 | 8.0 | 6.8 | 3.0 | Custom | Discovery, red teaming and runtime guardrails in one product you can host yourself |
| 12 | Orca Security AI-SPM Orca Security | 7.3 | 6.5 | 9.0 | 6.8 | 8.3 | 7.0 | 5.5 | Custom | A fast, agentless AI inventory and risk view across AWS, Azure and Google Cloud |
Expert reviews
Prisma AIRS is the most complete platform here for a security team that must map and defend all of an organisation's AI. It combines runtime protection, AI security posture management, model scanning, red teaming and an AI Agent Gateway, and it absorbed Protect AI when that deal closed on 22 July 2025.
At runtime it inspects prompts, responses and agent actions through an API (for apps you build) or through Palo Alto firewalls (network intercept). Prisma AIRS 3.0, launched on 23 March 2026, added posture checks for agents on 12 platforms, protection against tool misuse and memory manipulation, and a gateway that governs tool calls and model access. The August 2026 release added AI Discovery with Cortex AISPM, which inventories models, endpoints, datasets and agents in AWS, Azure and Google Cloud, plus hooks that check Claude and OpenAI Codex prompts before they run.
The docs are open about gaps: agent discovery tracks runtime activity on Amazon Bedrock but only reads configurations on Azure AI Foundry. Palo Alto says Gartner named it a Market Shaper among established vendors in its September 2026 quadrant for AI application security.
Pick it if you already run Palo Alto firewalls or Cortex and need one platform for many AI projects. Skip it if you protect one app or need prices up front.
What we like
- Runtime, posture, model scanning, red teaming and agent gateway in one platform
- Detailed public docs and monthly release notes
- Checks coding-agent prompts (Claude, Codex) before they run
- Works through an API or existing Palo Alto firewalls
Watch out for
- No public price list
- Best value inside the Palo Alto stack
- Agent runtime monitoring on Azure AI Foundry not yet supported
Wiz is the strongest AI-SPM product we found, and it now adds runtime detection. Google completed its $32 billion purchase on 11 March 2026; Wiz keeps its brand and still covers AWS, Azure, Google Cloud and Oracle Cloud.
The posture side is agentless. Connect your cloud accounts and Wiz builds an AI bill of materials of models, agents, tools, MCP servers, SDKs and AI services, then uses its security graph to show attack paths: for example, an internet-facing AI endpoint whose identity can read sensitive data. Built-in rules catch misconfigurations such as a SageMaker endpoint without encryption.
AI-APP, launched on 23 March 2026, watches three layers at runtime: model inputs and outputs, what agents and workloads execute, and cloud API and identity changes. Findings map to the OWASP Top 10 for LLM applications, and partners such as Cloudflare, TrojAI and Pillar Security feed in results. In April Wiz added AWS AgentCore, Copilot Studio, Salesforce Agentforce and Databricks.
The trade-off: Wiz's runtime layer is built for detection and investigation, and we found no in-line prompt blocking in its material. Its AWS Marketplace list prices are public, from $24,000 a year for 100 workloads.
Pick it if posture across several clouds is your main need. Skip it if you need an in-line prompt firewall.
What we like
- Agentless discovery across five clouds and SaaS AI
- Attack paths link AI assets to identities and data
- Runtime detection across model, workload and cloud layers
- Public AWS Marketplace list prices
Watch out for
- Runtime focus is detection, not in-line blocking
- AI features per pricing tier are not published
- Full value needs the wider Wiz platform
Zenity is the specialist to beat for securing AI agents, especially those staff build inside business software. It covers Microsoft 365 Copilot, Copilot Studio, Microsoft Foundry, Salesforce Agentforce, ServiceNow, ChatGPT Enterprise, Amazon Bedrock and Google Vertex AI, plus coding and personal agents such as Claude Code, Codex, Cursor and GitHub Copilot.
It splits the job in two. AI security posture management checks each agent's configuration, permissions, connected tools and data access. AI detection and response (AIDR) watches agents at runtime, mixes fixed rules with model-based checks of intent, and blocks risky actions in-line through native hooks and its MCP gateway. Detections map to OWASP and MITRE ATLAS. In August 2026 GitHub Copilot and Codex support became generally available.
The evidence is good for a startup. Zenity says Gartner named it one of two Market Shapers among nearly 20 startups in its September 2026 quadrant for AI application security, and it raised $125 million in August 2026. AWS added it as the AI security partner in Security Hub Extended in May 2026, and its listing shows rare public prices: $16 per million tokens for runtime protection, with a 30-day free trial.
Pick it if your agents live in Microsoft, Salesforce or ServiceNow tools. Skip it if you mainly need cloud infrastructure posture; Wiz or Orca fits better.
What we like
- Widest agent platform coverage we found
- In-line blocking through hooks and an MCP gateway
- Public usage prices and a 30-day trial on AWS
- Detections mapped to OWASP and MITRE ATLAS
Watch out for
- Little coverage of cloud infrastructure posture
- Vendor-reported results only
- Usage costs climb quickly at high token volumes
Noma covers the full agent life cycle and is the strongest choice for agents on employee laptops. It has four parts: AI-SPM that discovers agents, MCP servers and tools across cloud, SaaS and developer environments; access control over which agents may touch which data; multi-turn AI red teaming; and AI-DR, which follows a whole agent session (prompts, tool calls, data access and actions) rather than single messages.
On 28 September 2026 it extended this to employee machines, naming Claude Code, Cursor, Codex, Windsurf, Kiro, Antigravity and OpenClaw. Each agent, MCP server and skill lands in a registry as approved, needs review or blocked. Policies follow the person behind each agent and can allow reads widely while keeping create, update and delete rights for a smaller group. Noma says AI-DR catches prompt injection, data leakage, tool poisoning and agents drifting from their purpose.
Noma says Gartner named it a Market Shaper among startup vendors in September 2026, and claims deployment on more than 2.6 million endpoints. Prices are unusually visible on AWS Marketplace: $8 per million tokens scanned in a hybrid deployment or $16 as SaaS, or $200,000 a year for a full-platform contract.
Pick it if coding agents and MCP servers are spreading faster than you can track. Skip it if you want one tool for non-AI cloud posture too.
What we like
- Registry and tool-level control for agents on laptops
- Session-level runtime detection, not just single prompts
- Hybrid deployment option
- Public usage prices on AWS Marketplace
Watch out for
- Few published detection metrics
- Full-platform contract lists at $200,000 a year
- Not a general cloud security tool
Check Point's AI Defense Plane is the best pick when fast in-line blocking matters most. It is built on Lakera, the prompt-attack specialist whose purchase Check Point completed on 22 October 2025, plus agent discovery technology from Cyata, bought in early 2026.
Announced on 23 March 2026, it has three parts. Workforce AI Security governs staff use of AI tools. AI Agent Security inventories agents, the tools they reach and the actions they take, and blocks prompt injection, unauthorised tool calls, unsafe file access and runaway agent behaviour. AI Red Teaming tests prompts, workflows and tool use; it was in limited release at launch. Check Point says responses take under 50 milliseconds across more than 100 languages, and Lakera claimed detection above 98% with false positives below 0.5% when the deal was announced. Those are vendor figures.
Agent security can run as SaaS or self-hosted, which helps with data residency rules, and AWS Marketplace prices are public: $74,000 a year for 10 million guardrail requests and $100 per user a year for workforce protection.
The weak spot is posture. We found no cloud AI misconfiguration scanning comparable to Wiz or Orca.
Pick it if you need low-latency blocking you can host yourself. Skip it if AI-SPM across your clouds is the priority.
What we like
- Low-latency blocking with Lakera's prompt-attack engine
- SaaS or self-hosted deployment
- Public list prices on AWS Marketplace
- Covers employees, apps and agents in one console
Watch out for
- Light on cloud AI posture management
- Accuracy figures are vendor claims
- Red teaming was in limited release at launch
Cisco AI Defense is the most developer-friendly of the big-vendor platforms. It has four parts: AI Cloud Visibility finds models, agents and MCP servers in Amazon Bedrock, Google Vertex AI and Azure AI Foundry; supply chain scanning checks model files, repositories and MCP servers for malicious code; validation red-teams models with more than 200 attack techniques; and runtime protection screens prompts, responses and MCP traffic, with tool allowlists and blocklists.
The February 2026 expansion added an AI bill of materials, an MCP catalogue and real-time guardrails for agent actions, though Cisco said some features were still in development. For builders, the Agent Runtime Protection SDK wraps every LLM and MCP call with one line of code across LangChain, LangGraph, CrewAI, AutoGen and the three big cloud agent platforms, either calling the API directly or routing through Cisco's gateway. Findings map to NIST AI RMF, MITRE ATLAS and OWASP.
Explorer Edition gives free self-service red teaming, a useful way to judge the engine before a proof of concept. Runtime protection is sold in Essentials and Advantage tiers with no public price.
Pick it if you use Cisco Secure Access or want an SDK your developers will accept. Skip it if your agents mostly live in SaaS tools such as Copilot Studio.
What we like
- One-line agent SDK across major frameworks
- MCP inspection on requests and responses
- Detailed data sheet with framework mappings
- Free Explorer Edition for red teaming
Watch out for
- No public price for runtime protection
- Some announced features still in development
- Little SaaS agent coverage
Microsoft Defender for Cloud is the natural start for teams building on Azure, and the only big-vendor platform here with a public per-token price. It has two AI parts.
AI security posture management, in the Defender CSPM plan, discovers AI workloads in Azure OpenAI, Azure AI Foundry, Azure Machine Learning, Amazon Bedrock and Google Vertex AI, builds an AI bill of materials, runs attack path analysis and flags risky set-ups such as internet-facing AI endpoints with weak authentication. AI threat protection watches Azure-hosted models at runtime and raises alerts for jailbreaks, data leakage, data poisoning and credential theft, using Prompt Shields and Microsoft threat intelligence. Alerts land in Defender XDR next to your other incidents.
Microsoft's docs are clear about the limits. Runtime protection covers only Azure OpenAI and Azure AI model inference, scans text but not images or audio, and does not work on connected AWS accounts. Since 1 July 2026, agent-level security for Foundry and Copilot Studio agents needs a separate Agent 365 licence at $15 per user per month.
Prices are low and published: $0.0008 per 1,000 tokens scanned, with a 30-day trial capped at 75 billion tokens.
Pick it if your AI runs on Azure and you already use Defender. Skip it if your apps call models on AWS or Google Cloud.
What we like
- Published per-token price and a generous trial
- AI-SPM across Azure, AWS Bedrock and Vertex AI
- Alerts flow straight into Defender XDR
- Clear documentation of limits
Watch out for
- Runtime protection only for Azure-hosted models
- Agent security now needs an extra Agent 365 licence
- Alerts rather than in-line blocking for most threats
CrowdStrike brings AI detection and response to where it already lives: the endpoint. It bought Pangea, a prompt-security startup, in September 2025 and made Falcon AI Detection and Response (AIDR) generally available on 15 December 2025. CrowdStrike says AIDR inspects prompts, responses and agent actions, blocks more than 180 known prompt injection techniques, and stops credentials and regulated data from reaching AI tools.
On 1 September 2026 it added Falcon Guardian. The Falcon sensor finds known and shadow agents on Windows and Mac, including dormant ones, records who deployed them, and links each prompt to the tool calls and system actions that follow. Agents not on an approved list can be blocked, and alerts land in Falcon Next-Gen SIEM. An AI Gateway for central control was still in pre-beta at launch.
CrowdStrike claims 99% detection of prompt attacks at 100 milliseconds or less; Pangea's earlier figure of up to 99% at under 30 milliseconds came from its own internal benchmark. We found no independent test of either.
The gaps are posture and price: we found far less public detail on cloud AI posture than for Wiz or Microsoft, and no public price.
Pick it if you already run Falcon and worry about agents on staff machines. Skip it if your AI risk sits mostly in cloud services.
What we like
- Links prompts to what agents then do on the endpoint
- Finds dormant and shadow agents with the existing sensor
- Alerts in the same console as endpoint and identity threats
Watch out for
- No public pricing
- Little public detail on cloud AI posture
- AI Gateway not yet generally available at launch
SentinelOne's offer is strongest for controlling how staff and developers use AI. It completed its purchase of Prompt Security on 5 September 2025, and Prompt now sits in the Singularity console next to endpoint, cloud and identity tools.
Prompt Security works in the browser, desktop apps and code assistants (GitHub Copilot, Cursor and Claude Code), stopping secrets and proprietary code from leaking into prompts, and it protects homegrown AI apps with real-time runtime enforcement. SentinelOne says it recognises more than 15,000 AI tools and services. For agents it maps MCP servers, enforces least-privilege access and keeps searchable logs of agent actions. Prompt AI Agent Security, announced on 23 March 2026, adds real-time agent governance but was in preview at launch, alongside a new red-teaming service.
The posture side comes from Singularity Cloud Native Security, an agentless product whose AI-SPM inventories training jobs, notebooks, managed AI services and inference endpoints in AWS (SageMaker and Bedrock), Google Vertex AI and Azure, and traces attack paths. So SentinelOne covers both runtime and posture, but through two products.
We found no published detection metrics and no public price.
Pick it if you run SentinelOne and your biggest worry is data leaking through employee AI use. Skip it if you need mature agent runtime enforcement today.
What we like
- Strong control of employee AI use and code assistants
- MCP visibility and least-privilege controls for agents
- Agentless AI-SPM in the same platform
- One console with endpoint and cloud security
Watch out for
- No public pricing
- No published detection metrics
- Agent runtime product was still in preview
HiddenLayer is the independent choice for regulated firms that worry about the models themselves as well as prompts. It raised a $100 million Series B on 2 September 2026, led by Delta-v Capital, and says annual recurring revenue grew more than tenfold in a year.
Its platform has five modules: AI Discovery for shadow AI, supply chain security that checks models for tampering before you deploy them, attack simulation, runtime security, and the newest, Agent Harness Security. In March 2026 runtime security gained agent session replays, search across sessions for threat hunting, and real-time detection and blocking. It plugs into agent gateways and frameworks, so you can start with visibility and switch on blocking later. Agent Harness Security, launched on 3 August 2026, watches coding agents' prompts, tool calls, shell commands and file edits, and can redact secrets before the model sees them.
HiddenLayer says it added more than 50 platform customers in sectors such as banking, government, defence and healthcare, and it holds 39 granted patents. Cloud posture coverage is lighter than Wiz's, MCP is not named in its announcements, and there is no public price or free tier.
Pick it if model supply chain risk and coding agents both matter. Skip it if posture management is your first need.
What we like
- Covers model supply chain, testing and runtime
- Runtime protection for AI coding agents
- Independent and well funded
- Phased rollout from visibility to blocking
Watch out for
- No public pricing or free tier
- Lighter cloud AI posture coverage
- MCP coverage not spelled out
Pillar Security is the most complete of the young specialists, and one of few with a self-hosted option. Founded in 2024 in Miami and Tel Aviv, it links three jobs that others often sell separately.
Discovery catalogues agents, models, prompts, tools, MCP servers and skills across endpoints, code repositories, CI/CD pipelines, cloud and SaaS, and checks agent configurations, permissions and identities. RedGraph red-teams the whole system with multi-turn attacks, mapping how tools, permissions and data combine into attack paths. Runtime guardrails then block prompt injection, data leaks, tool misuse and unsafe commands before they run, and are tuned using what the red team found. Pillar runs as SaaS, inside your own cloud account or as a hybrid.
Recognition is growing. Pillar says Gartner named it a Pioneer among startup vendors in its September 2026 quadrant for AI application security and a 2026 Cool Vendor, and Wiz lists it as an integration partner for agent assessments. Pillar also maintains SAIL, an open framework that maps 91 risks across the agent life cycle.
The downsides: it is a small company, publishes no detection metrics and sells on quotes only.
Pick it if you want discovery, testing and runtime protection in one tool you can host yourself. Skip it if you prefer a large vendor's support and roadmap.
What we like
- Discovery, red teaming and runtime in one product
- SaaS, self-hosted or hybrid deployment
- Red-team findings tune the runtime guardrails
- Covers coding agents, MCP servers and skills
Watch out for
- Small, young company
- No published detection metrics
- No public pricing
Orca is the best pick for AI-SPM when you want answers on day one without installing anything. Its agentless SideScanning technology means connecting a cloud account is enough to find managed AI services, self-hosted and fine-tuned models, notebooks, datasets, vector stores and shadow AI. Orca says it covers Azure OpenAI, Amazon Bedrock and SageMaker, Google Vertex AI and more than 50 common AI models and packages such as PyTorch and Hugging Face, and it maps AI misconfigurations to the OWASP Top 10 for LLMs and for machine learning. It also flags sensitive data in training sets and links each AI risk to the identities and data around it.
Runtime coverage is newer. The optional Orca Sensor, announced in March 2026, records outbound LLM requests and inbound MCP activity, ties each to the workload and process that made it, and flags sensitive data leakage and prompt-injection risk. That is detection for the security team, not an in-line firewall that stops a bad prompt.
Orca offers a 30-day free trial through AWS Marketplace; contracts are quoted.
Pick it if you need a quick, agentless inventory of AI risk across AWS, Azure and Google Cloud. Skip it if you need to block attacks on agents in real time; pair it with a runtime tool.
What we like
- Agentless AI inventory in minutes
- Finds sensitive data in training sets
- Maps AI misconfigurations to OWASP
- 30-day free trial
Watch out for
- Runtime detection only, no in-line blocking
- Little agent-platform coverage
- No public price list
How we scored these tools
Each tool is scored 0–10 on the criteria below, using public evidence: independent benchmarks, vendor documentation and pricing pages, aggregate user ratings and reputable reviews. The overall score is the weighted average. Nobody pays to be listed. Read our full methodology.
| Criterion | Weight | What we look at |
|---|
| Runtime detection & response | 25% | Catches and blocks prompt injection (typed or hidden in content), data leakage, jailbreaks and tool misuse while apps and agents run, and supports investigation and response. |
| AI-SPM & discovery | 20% | Inventories models, agents, MCP servers, AI services and shadow AI, and flags misconfigurations, exposed endpoints and attack paths. |
| Agent & MCP coverage | 20% | Works across agent platforms (SaaS, cloud and coding agents on laptops), controls tool calls and MCP servers, and maps what each agent is allowed to do. |
| Deployment & integrations | 15% | Clouds and AI platforms supported, API, SDK, gateway and network options, self-hosting, and links to SIEM and SOC tools. |
| Evidence & transparency | 10% | Public documentation, stated limitations, mappings to OWASP and MITRE ATLAS, published metrics and independent recognition. |
| Pricing transparency | 10% | Published or marketplace list prices, free trials and how predictable costs are. |
What AI runtime security and monitoring is
An AI app or agent is most exposed once it is live. It reads untrusted content, talks to real users and holds real permissions. AI runtime security gives a security team eyes and controls on that live activity. The tools do four jobs:
- Discover. Find every model, agent, MCP server, AI service and AI tool in use, including shadow AI that nobody approved.
- Assess posture. Flag risky set-ups before an attack: an AI endpoint open to the internet, an agent with delete rights it never uses, or a training bucket full of customer data.
- Detect and block. Inspect prompts, responses and agent actions as they happen, and stop prompt injection, data leaks and tool misuse.
- Respond. Send alerts with the full session (prompt, tool calls, actions) to the security operations centre (SOC), and let analysts stop an agent or revoke its access.
Vendors collect this data in different ways: an API or SDK inside the app (Cisco, Prisma AIRS), a gateway that AI traffic passes through, hooks inside agent platforms and coding agents (Zenity, Noma), a sensor on servers or laptops (CrowdStrike, Orca), or read-only cloud APIs for posture (Wiz, Microsoft). Large organisations usually need two or more of these paths, so ask each vendor which ones it supports for your platforms. If agents and MCP are new to you, start with what is agentic AI and what is MCP.
Runtime security vs AI-SPM vs guardrails vs observability
These four categories overlap, and vendors use the labels loosely. The real difference is who runs the tool and how much of the organisation it covers.
| Category |
Main question |
Who runs it |
Scope |
Where to look |
| AI runtime security (AIDR) |
Is an attack happening now, and can we stop it? |
Security team, SOC |
All AI apps and agents |
This page |
| AI-SPM |
What AI do we have, and what is set up unsafely? |
Cloud security team |
All clouds and AI services |
This page |
| Guardrails |
Should this input or output be allowed? |
Developers |
One app at a time |
Best AI guardrails |
| LLM observability |
What did the app do, how good was it, what did it cost? |
Developers, ML engineers |
One app or team |
Best LLM observability tools |
Runtime protection vs AI-SPM. AI-SPM is preventive and mostly read-only: it scans configurations, permissions and data, often without installing anything. Runtime protection sits in the live path, or watches it, and acts on what is happening. AI-SPM tells you an agent could delete your customer records; runtime protection stops it when it tries.
Which tools do both? Prisma AIRS, Wiz, Zenity, Noma, Cisco and Pillar Security cover discovery, posture and runtime in one product. Microsoft and SentinelOne cover both through two products. Orca is posture-first with newer runtime detection, while Check Point, CrowdStrike and HiddenLayer are runtime-first with lighter posture features.
When you need which. One customer chatbot built by one team: guardrails plus observability are usually enough. Dozens of AI apps, agents built in Copilot Studio or Agentforce, coding agents on developer laptops, or an auditor asking for an AI inventory: you need runtime security and AI-SPM owned by the security team. Many organisations end up with all four layers.
Pricing guide (as of 9 October 2026)
Most of these platforms are sold on annual contracts with no public price. Where a list price exists, it is on the vendor's site, the Azure price list or AWS Marketplace. Treat marketplace list prices as a starting point: large contracts are usually negotiated.
| Tool |
Published price |
Unit |
Free option |
| Prisma AIRS |
Custom |
API licensed by monthly token volume |
None published |
| Wiz |
$24,000/yr Essential, $38,000/yr Advanced |
100 cloud workloads (AWS Marketplace) |
None published |
| Zenity |
$16 per million tokens (runtime), $2 per million (guardrails for AWS), $130 per resource a month (observability) |
AWS Marketplace |
30-day trial |
| Noma Security |
$8 (hybrid) or $16 (SaaS) per million tokens, $130 per asset, $650 per red-team test; or $200,000 a year |
AWS Marketplace |
None published |
| Check Point AI Defense Plane |
$74,000/yr for 10M guardrail requests; $100 per user a year for workforce protection |
AWS Marketplace |
None published |
| Cisco AI Defense |
Custom |
Essentials and Advantage tiers |
Explorer Edition (red teaming only) |
| Microsoft Defender for Cloud |
$0.0008 per 1,000 tokens; Defender CSPM $0.007 per resource-hour |
Tokens scanned; billable resources |
30 days, up to 75B tokens |
| CrowdStrike Falcon AIDR |
Custom |
Falcon platform |
None published |
| SentinelOne Prompt Security |
Custom |
Contract |
None published |
| HiddenLayer |
Custom |
Contract |
None published |
| Pillar Security |
Custom |
Contract |
None published |
| Orca Security |
Custom |
Contract |
30-day trial (AWS Marketplace) |
Worked example. Your agents send 500 million tokens a month through runtime inspection. Microsoft Defender for AI Services charges 500,000 x $0.0008 = $400 a month, but it only covers models hosted on Azure and mostly raises alerts rather than blocking in-line. Zenity's runtime protection costs 500 x $16 = $8,000 a month. Noma costs $4,000 a month in a hybrid deployment or $8,000 as SaaS. Check Point's guardrail tier works out at $0.0074 per request ($74,000 / 10 million), so 1 million requests a month costs about $7,400. If you need agent-level coverage for Foundry or Copilot Studio agents on Microsoft, add Agent 365 at $15 per user per month.
Expert tips- Run discovery in read-only mode before you buy blocking. The inventory usually turns up agents and MCP servers nobody registered, and it tells you which runtime path (SDK, gateway, platform hooks or endpoint sensor) you actually need.
- Start every runtime policy in monitor mode. Review two weeks of detections with the app owners, remove false positives, then switch only the high-confidence rules to block.
- Test indirect prompt injection through the channels your agents really read: hidden text in a PDF, a calendar invite, a support ticket and an MCP tool description. Attacks typed into the chat box are the easy part.
- Ask each vendor in writing whether its in-line service fails open or closed, where prompts are stored and for how long, and whether your data is used to train its models.
- Send alerts to your SIEM with session IDs and the agent's identity, then rehearse one response end to end: stop the agent session, revoke its token and review everything it touched.
Jargon explained
- AI-SPM
- AI security posture management: tools that list all the AI an organisation uses and flag risky set-ups, such as exposed endpoints or agents with too much access.
- AI-BOM
- AI bill of materials: an inventory of the models, datasets, libraries, agents and AI services inside your systems, like an ingredients list.
- AIDR
- AI detection and response: watching AI apps and agents at runtime for attacks and acting on them, the way endpoint detection and response does for laptops.
- MCP server
- A connector that gives an AI agent access to a tool or data source using the Model Context Protocol. A malicious or over-powered MCP server is a common attack path.
- Shadow AI
- AI tools, models or agents that staff use or build without approval from IT or security.
- Indirect prompt injection
- An attack hidden in content an AI reads, such as a web page, email or document, that tries to give the AI new instructions without the user noticing.
Frequently asked questions
What is AI runtime security?
AI runtime security protects AI apps, models and agents while they run in production. It inspects prompts, responses and agent actions, blocks or flags attacks such as prompt injection, data leakage and tool misuse, and sends alerts to the security team. Unlike guardrails, which developers add to one app, it covers all of an organisation's AI from one console.
What is AI-SPM (AI security posture management)?
AI-SPM finds every AI model, agent, MCP server and AI service your organisation uses and checks how each is set up. It flags risks such as internet-facing AI endpoints, agents with too many permissions, sensitive data in training sets and unapproved shadow AI. Wiz, Orca, Prisma AIRS (through Cortex AISPM) and Microsoft Defender CSPM are strong AI-SPM options.
What is the best AI runtime security tool in 2026?
For most large organisations, Palo Alto Networks Prisma AIRS, because it covers runtime protection, AI-SPM, model scanning, red teaming and agent control in one platform. Wiz is best for multicloud posture, Zenity and Noma Security for AI agents, and Microsoft Defender for Cloud for Azure teams that want published prices.
Do I need AI runtime security if I already use guardrails?
Not if you run one or two apps built by one team; guardrails plus monitoring may be enough. You need a runtime security platform once you have many AI apps, agents built in SaaS tools such as Copilot Studio or Agentforce, coding agents on laptops, or auditors asking for an AI inventory. The two work together; see our AI guardrails ranking.
How is AI runtime monitoring different from LLM observability?
Observability tools such as Langfuse and LangSmith record what an app did so developers can fix quality, speed and cost problems. Runtime security tools watch the same kind of traffic for attacks and policy breaches and send alerts to the SOC. See our LLM observability ranking for the developer side.
How much do AI runtime security tools cost?
Most are quote-only enterprise contracts. Published prices as of 9 October 2026 include Microsoft Defender for AI Services at $0.0008 per 1,000 tokens, Zenity at $16 per million tokens for runtime protection, Noma at $8 to $16 per million tokens, Check Point at $74,000 a year for 10 million guardrail requests, and Wiz from $24,000 a year for 100 cloud workloads (AWS Marketplace list prices).
Can these tools secure coding agents like Claude Code and Cursor?
Yes, several now do. Zenity covers Claude Code, Codex, Cursor and GitHub Copilot through hooks and an MCP gateway. Noma covers Claude Code, Cursor, Codex, Windsurf and others on employee machines. HiddenLayer's Agent Harness Security and CrowdStrike's Falcon Guardian watch agents at the endpoint, and Prisma AIRS checks Claude and Codex prompts before they run.
Do these tools stop prompt injection completely?
No. Language models cannot reliably tell instructions from data, and attackers adapt. Vendor claims of around 98% to 99% detection come from their own tests, and no independent comparative results were public as of October 2026. Combine runtime detection with least-privilege access, human approval for risky actions and regular red teaming.