Start here
AI security vendor landscape
Category names change every quarter; the problems don't. For each category: what buyers are solving, how products are deployed, the questions practitioners ask before a POC, and open-source tools that cover part of it.
Listing is not endorsement. No vendor pays to appear, and descriptions restate what each vendor says its product does. Missing a vendor or a correction? Tell the editor.
AI usage control (Shadow AI, GenAI DLP)
Employees paste data into public AI tools, copilots and agentic browsers the organisation never approved. Teams need to see which AI is used and stop sensitive data leaving in prompts and uploads.
Questions to ask before a POC
- Is this enforcement or only discovery? Show a prompt with a secret being blocked or redacted inline, not an alert afterwards.
- Which surfaces are covered today: web apps, desktop apps, IDE plugins, CLIs, agentic browsers, embedded AI inside SaaS?
- How do you handle personal accounts on sanctioned tools (personal ChatGPT vs the enterprise tenant)?
- What is the false-positive rate on our own prompts and files during the POC, and who tunes the classifiers?
- Where are prompts and responses stored, for how long, and in which region? Can logging of prompt content be turned off?
- Does it work on unmanaged / BYOD devices, and what breaks when TLS inspection is not possible?
- Can users get warn/justify flows instead of hard blocks, and are those decisions logged for audit?
How it is deployed
- SSE / SWG inline proxy: sees managed-device web traffic, needs TLS inspection, misses desktop apps that pin certificates
- Browser extension: prompt-level context and coaching, only covers browsers where it is installed
- Endpoint agent: covers desktop apps, CLIs and IDEs, adds another agent to manage
- API / CASB integration with sanctioned AI tenants: no inline blocking, after-the-fact visibility
Playbooks: Secure AI coding tools in your organisation, Get visibility into Shadow AI
Vendors 20
- Proofpoint AI Security Acuvity · part of Proofpoint · proofpoint.com Says it discovers shadow AI, governs agents by intent, protects data in AI interactions and secures MCP connections.
- Cato AI Security Aim Security · part of Cato Networks · catonetworks.com Cato says it inventories AI apps and agents, finds unapproved AI use, and applies inline guardrails against jailbreaks, prompt injection and off-policy output.
- Aurascape aurascape.ai Says it discovers AI usage including embedded AI, decodes prompt intent and enforces policy in real time, with guardrails for coding assistants.
- Cisco AI Defense Cisco · cisco.com Says it discovers AI workloads, validates models with algorithmic red teaming, applies runtime guardrails and controls access to third-party AI apps.
- Cloudflare AI Security Cloudflare · cloudflare.com Says it gives visibility into shadow AI, applies DLP to AI-bound traffic, and centralises MCP server access behind zero-trust policies and an AI gateway.
- Falcon AI Detection and Response (AIDR) CrowdStrike · crowdstrike.com Says it detects prompt injection and jailbreaks and blocks sensitive data, deployed via browser extensions, app SDKs, AI/API gateway integrations, an MCP proxy and cloud log analysis.
- Cyera cyera.com Data security platform that says it adds AI activity tracking, browser controls, AI-SPM, AI runtime protection and an agent guardian.
- Harmonic Security harmonic.security Says it detects shadow AI, prevents data leaks and coaches employees inline across browser, desktop and agentic workflows that use MCP.
- Knostic knostic.ai Says it secures coding assistants and desktop agents on the endpoint: shadow AI discovery, blocking unapproved MCP servers and vetting skills and extensions.
- Lakera part of Check Point · lakera.ai Says it screens AI app traffic for prompt attacks and data leakage (Lakera Guard), runs AI red teaming, and discovers workforce AI use.
- Akamai Workforce Protector LayerX · part of Akamai · akamai.com Browser extension that, per Akamai, discovers shadow AI, applies DLP to prompts, uploads and pastes, and governs AI agent actions in the browser.
- Microsoft Purview (DSPM for AI) Microsoft · learn.microsoft.com Microsoft documents sensitivity labels, DLP, insider-risk and audit controls for Microsoft 365 Copilot and other generative AI apps.
- Netskope One AI Security Netskope · netskope.com Says it discovers AI tools and activity, protects AI interactions with DLP, and offers GenAI app security, an agentic broker, AI gateway, guardrails, agent action control and AI red teaming.
- Nightfall nightfall.ai Says it controls how data moves across AI agents, MCP servers, endpoints and SaaS, including shadow AI prevention and an MCP gateway.
- Pillar Security pillar.security Says it discovers and governs AI agents and workflows, maps the AI attack surface, red-teams AI apps and applies runtime guardrails.
- Prompt Security part of SentinelOne · prompt.security SentinelOne says it discovers employee and shadow AI use, redacts data sent to AI code assistants, protects homegrown AI apps and governs MCP servers.
- Reco reco.ai Says it finds AI agents across SaaS apps, maps their owners and access, monitors agent runtime activity and detects shadow AI.
- Singulr AI singulr.ai Says it discovers and governs public AI tools, embedded AI and in-house agents, and extends runtime governance to endpoint agents such as Claude Code and Cursor.
- WitnessAI witness.ai Says it gives network-level visibility of employee and agent AI use, intent-based controls, MCP server allowlisting and runtime guardrails.
- Zscaler AI Security Zscaler · zscaler.com Says it sits inline across AI traffic, discovers models, agents, MCP servers and shadow AI, applies inline DLP, and brokers MCP and A2A traffic through an AI gateway.
AI runtime protection (LLM firewalls, guardrails)
Your own AI apps and agents take untrusted input (users, documents, web pages, tool output). Teams want inline checks for prompt injection, data leakage and unsafe output before they reach the model, the user or a tool.
Questions to ask before a POC
- What exactly is blocked inline versus only flagged? Run our own indirect-injection test set through it.
- What latency does it add at p50 and p99, including on streamed responses?
- What is the false-positive rate on our real traffic, and can we tune or allowlist per app?
- Does it inspect tool calls, tool outputs and retrieved documents, or only user prompts and final answers?
- Does it handle multi-turn attacks and non-English or encoded payloads?
- Can it be self-hosted or run in our region, and are prompts retained or used to train the vendor's detectors?
- What happens on timeout or outage: fail open or fail closed, and is that configurable?
How it is deployed
- Inline AI gateway / reverse proxy: one choke point for all model traffic, adds latency and a new dependency
- SDK / API call from the app: precise context per request, needs code changes in every app
- Integration with an existing API or AI gateway (LiteLLM, Kong, cloud gateways): reuses routing, coverage depends on the gateway
- Cloud-native service in the model provider's platform: easy if you are single-cloud, harder across providers
Open source that covers part of it
LLM Guard, NeMo Guardrails, Guardrails AI, PurpleLlama (Llama Guard, Prompt Guard, LlamaFirewall), LangKit
Techniques: LLM01 · LLM02 · LLM05 · LLM07 · LLM10 · ASI01
Playbooks: Red-team an LLM application, Secure RAG and vector stores
Vendors 21
- Cato AI Security Aim Security · part of Cato Networks · catonetworks.com Cato says it inventories AI apps and agents, finds unapproved AI use, and applies inline guardrails against jailbreaks, prompt injection and off-policy output.
- F5 AI Guardrails, F5 AI Red Team CalypsoAI · part of F5 · f5.com F5 says AI Guardrails protects generative AI apps against adversarial attacks and data leakage; AI Red Team tests them. calypsoai.com now redirects to F5.
- Cisco AI Defense Cisco · cisco.com Says it discovers AI workloads, validates models with algorithmic red teaming, applies runtime guardrails and controls access to third-party AI apps.
- Falcon AI Detection and Response (AIDR) CrowdStrike · crowdstrike.com Says it detects prompt injection and jailbreaks and blocks sensitive data, deployed via browser extensions, app SDKs, AI/API gateway integrations, an MCP proxy and cloud log analysis.
- Model Armor Google Cloud · cloud.google.com Says it screens prompts, responses and agent interactions for prompt injection, jailbreaks, sensitive data and malicious URLs, model- and cloud-agnostic.
- HiddenLayer hiddenlayer.com Says it offers AI discovery, model scanning for malicious or backdoored models, red teaming, guardrails and protection for agentic and MCP systems.
- Lakera part of Check Point · lakera.ai Says it screens AI app traffic for prompt attacks and data leakage (Lakera Guard), runs AI red teaming, and discovers workforce AI use.
- Lasso Security lasso.security Says it offers AI discovery and AI-BOM, posture management, automated red teaming, runtime enforcement and AI detection and response.
- Netskope One AI Security Netskope · netskope.com Says it discovers AI tools and activity, protects AI interactions with DLP, and offers GenAI app security, an agentic broker, AI gateway, guardrails, agent action control and AI red teaming.
- NeuralTrust neuraltrust.ai Says it offers agent runtime security (TrustGuard) and an agent gateway (TrustGate) that connects agents to models, MCPs and APIs with identity-aware controls.
- Operant AI operant.ai Says it discovers and defends AI apps, agents and MCP in real time with a runtime firewall, MCP gateway, endpoint protection for coding agents and red teaming.
- Prisma AIRS Palo Alto Networks · paloaltonetworks.com Says it covers AI gateway, runtime security, red teaming, posture management, AI supply-chain (model) security, agent identity and coding-agent security; protectai.com (Protect AI) now redirects here.
- Pangea part of CrowdStrike · pangea.cloud Says it offers security guardrails for AI applications that developers add with a few lines of code.
- Pillar Security pillar.security Says it discovers and governs AI agents and workflows, maps the AI attack surface, red-teams AI apps and applies runtime guardrails.
- Prompt Security part of SentinelOne · prompt.security SentinelOne says it discovers employee and shadow AI use, redacts data sent to AI code assistants, protects homegrown AI apps and governs MCP servers.
- Repello AI repello.ai Says it inventories AI systems, runs autonomous red teaming across LLMs, agents and MCP, and enforces runtime guardrails.
- SPLX part of Zscaler · splx.ai Says it provides automated AI red teaming, runtime input/output guardrails and AI asset discovery with an AI-BOM.
- Straiker straiker.ai Says it provides agent visibility and governance, adversarial testing and runtime defense, including cutting tool access or taking an agent offline.
- TrojAI part of A10 Networks · troj.ai Says it secures AI agents from build time to runtime with AI red teaming, model guardrails and agent runtime protection.
- FortiAIGate Virtue AI · part of Fortinet · fortinet.com Fortinet says FortiAIGate is a self-hosted AI gateway that protects LLMs, agents and MCP tools, enforces usage policy and caps spend; virtueai.com now redirects here.
- WitnessAI witness.ai Says it gives network-level visibility of employee and agent AI use, intent-based controls, MCP server allowlisting and runtime guardrails.
AI red teaming and testing
Teams need repeatable adversarial testing of LLM apps and agents before and after release, with findings that show real impact rather than jailbreak screenshots.
Questions to ask before a POC
- Does it test agents with tools, memory and MCP servers, or only a chat endpoint?
- How are findings scored: by business impact (data access, tool misuse) or by policy-violating text?
- Can we add our own attack datasets and app-specific policies, and re-run them as regression tests in CI?
- How does it test indirect prompt injection via documents, web pages and tool output?
- How much of the attack library is public (garak, PyRIT, promptfoo) versus proprietary, and how often is it updated?
- What does a finding include: reproducible transcript, root cause, suggested fix?
- Where do test prompts and model responses go, and can the engine run against private models?
How it is deployed
- SaaS scanner against your endpoint: quick to start, needs network access to the target
- CLI / CI integration: runs on every change, needs engineering ownership
- Self-hosted engine: keeps prompts and responses in-house, more to operate
- Expert-led service: deeper manual work, point-in-time
Open source that covers part of it
garak, PyRIT, promptfoo, Giskard, AgentDojo, Adversarial Robustness Toolbox
Techniques: LLM01 · LLM02 · LLM05 · LLM06 · LLM07 · LLM09 · ASI01 · ASI02
Playbooks: Red-team an LLM application, Secure RAG and vector stores, Prepare for AI-powered attackers
Vendors 19
- Adversa AI adversa.ai Says it monitors coding agents and stops harmful actions, and offers red teaming for LLMs, agents and MCP.
- F5 AI Guardrails, F5 AI Red Team CalypsoAI · part of F5 · f5.com F5 says AI Guardrails protects generative AI apps against adversarial attacks and data leakage; AI Red Team tests them. calypsoai.com now redirects to F5.
- Capsule Security capsulesecurity.io Says it discovers agents across homegrown, SaaS and endpoint environments and stops rogue agent actions at runtime based on intent, with whitebox red teaming.
- Cisco AI Defense Cisco · cisco.com Says it discovers AI workloads, validates models with algorithmic red teaming, applies runtime guardrails and controls access to third-party AI apps.
- Giskard giskard.ai Says it provides continuous red teaming of AI agents, with data-residency options and managed remediation.
- HiddenLayer hiddenlayer.com Says it offers AI discovery, model scanning for malicious or backdoored models, red teaming, guardrails and protection for agentic and MCP systems.
- Lakera part of Check Point · lakera.ai Says it screens AI app traffic for prompt attacks and data leakage (Lakera Guard), runs AI red teaming, and discovers workforce AI use.
- Lasso Security lasso.security Says it offers AI discovery and AI-BOM, posture management, automated red teaming, runtime enforcement and AI detection and response.
- Mindgard mindgard.ai Says it performs attacker-style reconnaissance and continuous automated red teaming of AI models, agents and applications.
- Netskope One AI Security Netskope · netskope.com Says it discovers AI tools and activity, protects AI interactions with DLP, and offers GenAI app security, an agentic broker, AI gateway, guardrails, agent action control and AI red teaming.
- Noma Security noma.security Says it discovers agents, MCP servers, skills and tools, controls agent behaviour and access at runtime, and red-teams agents.
- Operant AI operant.ai Says it discovers and defends AI apps, agents and MCP in real time with a runtime firewall, MCP gateway, endpoint protection for coding agents and red teaming.
- Prisma AIRS Palo Alto Networks · paloaltonetworks.com Says it covers AI gateway, runtime security, red teaming, posture management, AI supply-chain (model) security, agent identity and coding-agent security; protectai.com (Protect AI) now redirects here.
- Pillar Security pillar.security Says it discovers and governs AI agents and workflows, maps the AI attack surface, red-teams AI apps and applies runtime guardrails.
- Promptfoo part of OpenAI · promptfoo.dev Open-source-based AI security testing and red teaming for LLM apps and agents in the development workflow; its site says it is now part of OpenAI.
- Repello AI repello.ai Says it inventories AI systems, runs autonomous red teaming across LLMs, agents and MCP, and enforces runtime guardrails.
- SPLX part of Zscaler · splx.ai Says it provides automated AI red teaming, runtime input/output guardrails and AI asset discovery with an AI-BOM.
- Straiker straiker.ai Says it provides agent visibility and governance, adversarial testing and runtime defense, including cutting tool access or taking an agent offline.
- TrojAI part of A10 Networks · troj.ai Says it secures AI agents from build time to runtime with AI red teaming, model guardrails and agent runtime protection.
AI agent security (inventory, policy, runtime)
Agents built on SaaS platforms, in-house frameworks and on endpoints act with real permissions. Teams need to know which agents exist, what they can reach, and to stop unsafe actions while they run.
Questions to ask before a POC
- Discovery without enforcement is just inventory: which actions can it block or pause in real time, and on which platforms?
- Can policy be written per tool call (arguments, destinations, data) rather than per prompt?
- Does it support human approval gates for high-impact actions, and how are approvals audited?
- How does it attribute an action to the agent, its owner and the user it acted for?
- Coverage: SaaS agent platforms, in-house frameworks, MCP clients, local and coding agents — which are GA today?
- What does 'no proxy, no SDK' mean in practice: which signals come from logs after the fact?
- What is the detection latency and false-positive rate on our agents' normal behaviour?
- Can it kill or isolate a misbehaving agent, and how quickly?
How it is deployed
- API connectors to agent platforms (Copilot Studio, Agentforce, Bedrock…): fast inventory and posture, limited inline control
- SDK / framework hooks in your agents: per-tool-call policy, needs code changes
- Gateway in front of models and tools: central enforcement, only sees traffic routed through it
- Endpoint agent for local and coding agents: sees processes and files, endpoint rollout required
Open source that covers part of it
Agentic Radar, PurpleLlama (Llama Guard, Prompt Guard, LlamaFirewall), AgentDojo
Techniques: LLM06 · ASI01 · ASI02 · ASI05 · ASI06 · ASI07 · ASI08 · ASI10
Playbooks: Lock down MCP servers and clients, Give AI agents identity and least privilege, Respond to a compromised AI agent
Vendors 22
- Proofpoint AI Security Acuvity · part of Proofpoint · proofpoint.com Says it discovers shadow AI, governs agents by intent, protects data in AI interactions and secures MCP connections.
- Astrix Security part of Cisco · astrix.security Says it inventories AI agents, MCP servers and non-human identities and enforces least-privilege access with audit trails.
- Capsule Security capsulesecurity.io Says it discovers agents across homegrown, SaaS and endpoint environments and stops rogue agent actions at runtime based on intent, with whitebox red teaming.
- Cyera cyera.com Data security platform that says it adds AI activity tracking, browser controls, AI-SPM, AI runtime protection and an agent guardian.
- HiddenLayer hiddenlayer.com Says it offers AI discovery, model scanning for malicious or backdoored models, red teaming, guardrails and protection for agentic and MCP systems.
- Invariant Labs part of Snyk · invariantlabs.ai Makes MCP Scan (MCP server scanning), Guardrails (a contextual security layer for agents) and Explorer for agent traces.
- Lakera part of Check Point · lakera.ai Says it screens AI app traffic for prompt attacks and data leakage (Lakera Guard), runs AI red teaming, and discovers workforce AI use.
- Lasso Security lasso.security Says it offers AI discovery and AI-BOM, posture management, automated red teaming, runtime enforcement and AI detection and response.
- Netskope One AI Security Netskope · netskope.com Says it discovers AI tools and activity, protects AI interactions with DLP, and offers GenAI app security, an agentic broker, AI gateway, guardrails, agent action control and AI red teaming.
- NeuralTrust neuraltrust.ai Says it offers agent runtime security (TrustGuard) and an agent gateway (TrustGate) that connects agents to models, MCPs and APIs with identity-aware controls.
- Nightfall nightfall.ai Says it controls how data moves across AI agents, MCP servers, endpoints and SaaS, including shadow AI prevention and an MCP gateway.
- Noma Security noma.security Says it discovers agents, MCP servers, skills and tools, controls agent behaviour and access at runtime, and red-teams agents.
- Prisma AIRS Palo Alto Networks · paloaltonetworks.com Says it covers AI gateway, runtime security, red teaming, posture management, AI supply-chain (model) security, agent identity and coding-agent security; protectai.com (Protect AI) now redirects here.
- Pillar Security pillar.security Says it discovers and governs AI agents and workflows, maps the AI attack surface, red-teams AI apps and applies runtime guardrails.
- Reco reco.ai Says it finds AI agents across SaaS apps, maps their owners and access, monitors agent runtime activity and detects shadow AI.
- Runlayer runlayer.com Says it discovers agents, MCP servers and skills in use and governs MCP and LLM calls through one gateway with runtime defenses.
- Singulr AI singulr.ai Says it discovers and governs public AI tools, embedded AI and in-house agents, and extends runtime governance to endpoint agents such as Claude Code and Cursor.
- Straiker straiker.ai Says it provides agent visibility and governance, adversarial testing and runtime defense, including cutting tool access or taking an agent offline.
- Token Security token.security Says it discovers and governs AI agents and non-human identities, including access right-sizing, lifecycle management and threat detection.
- TrojAI part of A10 Networks · troj.ai Says it secures AI agents from build time to runtime with AI red teaming, model guardrails and agent runtime protection.
- Zenity zenity.io Says it discovers AI agents across SaaS, cloud and endpoints, manages posture, and evaluates agent actions at runtime to enforce what they may do.
- Zscaler AI Security Zscaler · zscaler.com Says it sits inline across AI traffic, discovers models, agents, MCP servers and shadow AI, applies inline DLP, and brokers MCP and A2A traffic through an AI gateway.
MCP security (gateways, scanning, registries)
MCP servers give agents tools, data and credentials, and tool descriptions can carry hidden instructions. Teams want to know which servers are in use, allow only vetted ones, and control what flows through them.
Questions to ask before a POC
- How does it find MCP servers that developers configured locally (stdio) and never routed through the gateway?
- Does it detect tool poisoning and rug pulls, such as tool descriptions that change after approval or contain invisible Unicode?
- Can policy be set per tool and per argument, not just per server?
- How are downstream credentials handled: does the gateway hold tokens, or do agents still see raw secrets?
- Does it support both stdio and remote (HTTP) transports, and OAuth for remote servers?
- What latency does the gateway add, and what happens to agents when it is down?
- Is there a curated or scanned registry, and how often are listed servers re-checked?
How it is deployed
- MCP gateway / proxy with a private registry: central allowlist and logging, clients must be pointed at it
- Endpoint discovery of MCP configs: finds shadow servers on laptops, enforcement depends on the endpoint agent
- Static scanner for MCP servers and tool descriptions: cheap pre-install check, no runtime protection
- OpenTelemetry / log-based visibility: no traffic rerouting, detection only
Open source that covers part of it
Techniques: ASI04 · ASI02 · LLM03 · LLM01
Playbooks: Secure AI coding tools in your organisation, Lock down MCP servers and clients, Respond to a compromised AI agent
Vendors 21
- Proofpoint AI Security Acuvity · part of Proofpoint · proofpoint.com Says it discovers shadow AI, governs agents by intent, protects data in AI interactions and secures MCP connections.
- Aembit aembit.io Says it gives agents and workloads runtime IAM: scoped credentials kept out of agents, access to MCP servers by policy, audit and a kill switch.
- Cloudflare AI Security Cloudflare · cloudflare.com Says it gives visibility into shadow AI, applies DLP to AI-bound traffic, and centralises MCP server access behind zero-trust policies and an AI gateway.
- Falcon AI Detection and Response (AIDR) CrowdStrike · crowdstrike.com Says it detects prompt injection and jailbreaks and blocks sensitive data, deployed via browser extensions, app SDKs, AI/API gateway integrations, an MCP proxy and cloud log analysis.
- Harmonic Security harmonic.security Says it detects shadow AI, prevents data leaks and coaches employees inline across browser, desktop and agentic workflows that use MCP.
- Invariant Labs part of Snyk · invariantlabs.ai Makes MCP Scan (MCP server scanning), Guardrails (a contextual security layer for agents) and Explorer for agent traces.
- Keycard keycard.ai Says it gives each agent its own identity and issues just-in-time scoped access to MCP servers and APIs, on behalf of users, under runtime policy.
- Knostic knostic.ai Says it secures coding assistants and desktop agents on the endpoint: shadow AI discovery, blocking unapproved MCP servers and vetting skills and extensions.
- MCP Manager part of Usercentrics · mcpmanager.ai MCP gateway that says it provides a private MCP registry, logging, PII filtering, RBAC tool provisioning and OAuth enforcement.
- Netskope One AI Security Netskope · netskope.com Says it discovers AI tools and activity, protects AI interactions with DLP, and offers GenAI app security, an agentic broker, AI gateway, guardrails, agent action control and AI red teaming.
- NeuralTrust neuraltrust.ai Says it offers agent runtime security (TrustGuard) and an agent gateway (TrustGate) that connects agents to models, MCPs and APIs with identity-aware controls.
- Nightfall nightfall.ai Says it controls how data moves across AI agents, MCP servers, endpoints and SaaS, including shadow AI prevention and an MCP gateway.
- Noma Security noma.security Says it discovers agents, MCP servers, skills and tools, controls agent behaviour and access at runtime, and red-teams agents.
- Operant AI operant.ai Says it discovers and defends AI apps, agents and MCP in real time with a runtime firewall, MCP gateway, endpoint protection for coding agents and red teaming.
- Prompt Security part of SentinelOne · prompt.security SentinelOne says it discovers employee and shadow AI use, redacts data sent to AI code assistants, protects homegrown AI apps and governs MCP servers.
- Runlayer runlayer.com Says it discovers agents, MCP servers and skills in use and governs MCP and LLM calls through one gateway with runtime defenses.
- Unbound getunbound.ai Says it inventories AI coding agents and enforces policy on their terminal commands, MCP connections and data access.
- FortiAIGate Virtue AI · part of Fortinet · fortinet.com Fortinet says FortiAIGate is a self-hosted AI gateway that protects LLMs, agents and MCP tools, enforces usage policy and caps spend; virtueai.com now redirects here.
- WitnessAI witness.ai Says it gives network-level visibility of employee and agent AI use, intent-based controls, MCP server allowlisting and runtime guardrails.
- Zenity zenity.io Says it discovers AI agents across SaaS, cloud and endpoints, manages posture, and evaluates agent actions at runtime to enforce what they may do.
- Zscaler AI Security Zscaler · zscaler.com Says it sits inline across AI traffic, discovers models, agents, MCP servers and shadow AI, applies inline DLP, and brokers MCP and A2A traffic through an AI gateway.
Agent identity and authorization
Agents often run on shared service accounts or inherit everything the user can do. Teams want each agent to have its own identity, short-lived scoped credentials, and an audit trail of who it acted for.
Questions to ask before a POC
- How does it model delegation: agent acting as itself vs on behalf of a user (for example OAuth token exchange, RFC 8693)?
- Are credentials short-lived and scoped per task, and can the agent ever see the raw secret?
- Can it discover existing agents, service accounts and API keys, or only govern new ones?
- Does it authorize individual tool and MCP calls at runtime, or only grant access at login?
- How are agent identities tied to an accountable human owner, and what happens when the owner leaves?
- Can access be revoked instantly across all agents (kill switch), and how is that tested?
- How does it prevent a local agent or process from spoofing another agent's identity?
How it is deployed
- Identity provider extension (agent identities in the IdP): fits existing SSO and governance, tied to that IdP
- Runtime authorization service / credential broker: issues scoped tokens per request, sits in the request path
- NHI discovery and governance via API: finds service accounts, keys and agents, does not issue credentials itself
- Developer SDKs for agent and MCP auth (OAuth, token exchange): precise, requires app changes
Techniques: ASI03 · LLM06 · ASI07
Playbooks: Give AI agents identity and least privilege, Respond to a compromised AI agent
Vendors 10
- Aembit aembit.io Says it gives agents and workloads runtime IAM: scoped credentials kept out of agents, access to MCP servers by policy, audit and a kill switch.
- Astrix Security part of Cisco · astrix.security Says it inventories AI agents, MCP servers and non-human identities and enforces least-privilege access with audit trails.
- Clutch Security clutch.security Says it discovers non-human identities, AI agents and secrets, maps them to owners, and sets boundaries on agent permissions and credentials.
- Descope descope.com Customer and agentic identity platform that says it provides auth for AI agents and MCP servers through workflows, SDKs and APIs.
- Keycard keycard.ai Says it gives each agent its own identity and issues just-in-time scoped access to MCP servers and APIs, on behalf of users, under runtime policy.
- Microsoft Entra Agent ID Microsoft · learn.microsoft.com Microsoft documents agent identities in Entra for managing, governing lifecycles and protecting agent access to resources with Zero Trust policies.
- Oasis Security part of Cyera · oasis.security Says it provides agentic access management: mapping agents, identities and permissions, provisioning agent identities and enforcing policy.
- Okta for AI Agents, Auth0 for AI Agents Okta · okta.com Says it gives every agent its own identity, discovers agents and enforces policy at tool calls, for workforce (Okta) and customer-facing (Auth0) agents.
- Prisma AIRS Palo Alto Networks · paloaltonetworks.com Says it covers AI gateway, runtime security, red teaming, posture management, AI supply-chain (model) security, agent identity and coding-agent security; protectai.com (Protect AI) now redirects here.
- Token Security token.security Says it discovers and governs AI agents and non-human identities, including access right-sizing, lifecycle management and threat detection.
AI-SPM, AI-BOM and model supply chain
Models, datasets, AI services, notebooks and agent components spread across clouds and repos. Teams need an inventory (AI-BOM), misconfiguration and exposure findings, and scanning of model files before they load.
Questions to ask before a POC
- What does the AI-BOM include: models, datasets, AI services, agents, MCP servers, prompts?
- Which model formats are scanned (pickle, safetensors, GGUF, ONNX…), and are findings explained or just flagged?
- How are findings prioritised: exposure and data access paths, or a flat list of misconfigurations?
- Does it cover self-hosted models and GPU workloads, not only managed AI services?
- Can it block an unsafe model from being pulled or deployed, or only report it?
- How often is inventory refreshed, and how does it handle multiple clouds and on-prem?
How it is deployed
- Agentless cloud API scanning: broad inventory quickly, no runtime view
- Model file / artifact scanning in CI or registries: catches malicious pickles and backdoors, only on scanned paths
- Code and repo scanning for AI SDK usage: finds shadow AI in code, misses runtime-only usage
Open source that covers part of it
ModelScan, Fickling, picklescan, Agentic Radar
Techniques: LLM03 · LLM04 · LLM08 · ASI04
Playbooks: Secure RAG and vector stores, Get visibility into Shadow AI
Vendors 12
- Cisco AI Defense Cisco · cisco.com Says it discovers AI workloads, validates models with algorithmic red teaming, applies runtime guardrails and controls access to third-party AI apps.
- Cyera cyera.com Data security platform that says it adds AI activity tracking, browser controls, AI-SPM, AI runtime protection and an agent guardian.
- HiddenLayer hiddenlayer.com Says it offers AI discovery, model scanning for malicious or backdoored models, red teaming, guardrails and protection for agentic and MCP systems.
- Lasso Security lasso.security Says it offers AI discovery and AI-BOM, posture management, automated red teaming, runtime enforcement and AI detection and response.
- Noma Security noma.security Says it discovers agents, MCP servers, skills and tools, controls agent behaviour and access at runtime, and red-teams agents.
- Prisma AIRS Palo Alto Networks · paloaltonetworks.com Says it covers AI gateway, runtime security, red teaming, posture management, AI supply-chain (model) security, agent identity and coding-agent security; protectai.com (Protect AI) now redirects here.
- Pillar Security pillar.security Says it discovers and governs AI agents and workflows, maps the AI attack surface, red-teams AI apps and applies runtime guardrails.
- Repello AI repello.ai Says it inventories AI systems, runs autonomous red teaming across LLMs, agents and MCP, and enforces runtime guardrails.
- SPLX part of Zscaler · splx.ai Says it provides automated AI red teaming, runtime input/output guardrails and AI asset discovery with an AI-BOM.
- Wiz AI-SPM Wiz · part of Google · wiz.io Says it builds an agentless AI-BOM, detects AI misconfigurations and exposed AI endpoints, and extends attack-path analysis to AI.
- Zenity zenity.io Says it discovers AI agents across SaaS, cloud and endpoints, manages posture, and evaluates agent actions at runtime to enforce what they may do.
- Zscaler AI Security Zscaler · zscaler.com Says it sits inline across AI traffic, discovers models, agents, MCP servers and shadow AI, applies inline DLP, and brokers MCP and A2A traffic through an AI gateway.
AI SOC / agentic SOC
SOC teams cannot investigate every alert. AI SOC products use LLM agents to triage and investigate alerts, but practitioners question accuracy and whether it is just AI added to SOAR.
Questions to ask before a POC
- On our historical alerts with known outcomes, what share of true and false positives did it classify correctly?
- Does every verdict show the evidence and queries it ran, so an analyst can verify it?
- Which actions can it take on its own (containment, account disable), and what approval gates exist?
- What read and write permissions does it need in our tools, and how are those credentials stored?
- How does it handle alert sources and log types it has not seen before?
- Where is our telemetry processed and stored, and is it used to train shared models?
- How is pricing measured (alerts, investigations, seats), and what happens at alert spikes?
How it is deployed
- SaaS connected to SIEM/EDR/IdP via API: fast start, read access to many tools
- Self-hosted or sovereign deployment: data stays in-house, more operations
- Co-managed / MDR service built on the platform: less staff needed, less control
Playbooks: Evaluate an AI SOC or agentic triage tool, Prepare for AI-powered attackers
Vendors 6
- 7AI 7ai.com Says its AI agents detect, investigate, respond and hunt across the SOC stack with humans on the loop.
- Dropzone AI dropzone.ai Says its AI agents investigate alerts, hunt threats and respond, running on the security stack you already have.
- Exaforce exaforce.com Says it offers an agentic SOC platform and MDR with AI agents for detection, triage, investigation and hunting.
- Microsoft Security Copilot Microsoft · microsoft.com Says its agents work across Defender, Entra, Intune and Purview to help detect, investigate and respond to incidents.
- Prophet Security prophetsecurity.ai Says its AI SOC agents automate alert triage, investigation, response, threat hunting and detection engineering.
- Torq torq.io Says it provides an AI SOC platform with agents for triage, investigation, response and threat hunting.
AI coding agent security
Coding agents (Claude Code, Cursor, Copilot, Codex…) run with developer permissions, terminal access, MCP servers and skills on laptops. Teams want to see which agents run where and stop destructive commands, secret exposure and unvetted extensions.
Questions to ask before a POC
- Which coding agents and IDEs are supported today, and on which OSes?
- Can it block specific commands (rm -rf, force push, prod DB writes) before they run, or only log them?
- How does it stop secrets in files, env vars and transcripts from reaching the model?
- Does it vet MCP servers, skills and extensions before install, and detect changes later?
- What does a developer experience when something is blocked, and can they request an exception?
- What does it collect from developer machines (code, prompts, transcripts), and where is it stored?
How it is deployed
- Endpoint agent / local policy engine: sees commands and files, needs rollout to developer machines
- Managed settings and hooks pushed to the tools: low overhead, limited to tools that support them
- Gateway for model and MCP traffic: central control, local actions are invisible to it
Open source that covers part of it
Techniques: ASI05 · ASI04 · ASI02 · ASI03 · LLM02
Playbooks: Secure AI coding tools in your organisation
Vendors 8
- Adversa AI adversa.ai Says it monitors coding agents and stops harmful actions, and offers red teaming for LLMs, agents and MCP.
- Aurascape aurascape.ai Says it discovers AI usage including embedded AI, decodes prompt intent and enforces policy in real time, with guardrails for coding assistants.
- Knostic knostic.ai Says it secures coding assistants and desktop agents on the endpoint: shadow AI discovery, blocking unapproved MCP servers and vetting skills and extensions.
- Operant AI operant.ai Says it discovers and defends AI apps, agents and MCP in real time with a runtime firewall, MCP gateway, endpoint protection for coding agents and red teaming.
- Prisma AIRS Palo Alto Networks · paloaltonetworks.com Says it covers AI gateway, runtime security, red teaming, posture management, AI supply-chain (model) security, agent identity and coding-agent security; protectai.com (Protect AI) now redirects here.
- Prompt Security part of SentinelOne · prompt.security SentinelOne says it discovers employee and shadow AI use, redacts data sent to AI code assistants, protects homegrown AI apps and governs MCP servers.
- Singulr AI singulr.ai Says it discovers and governs public AI tools, embedded AI and in-house agents, and extends runtime governance to endpoint agents such as Claude Code and Cursor.
- Unbound getunbound.ai Says it inventories AI coding agents and enforces policy on their terminal commands, MCP connections and data access.
Reviewed 2026-11-09