Out Of AI℠ — All Things AI
Start here ↗
The business protection desk

AI Security

Protect your business,
data and AI agents.

Find controls for the AI your company uses—and for attacks that use AI against your company. Start with the risk, then build a shortlist.

01 / Match the protection to the problem

What do you need to protect?

See a business example ↓

Showing all 8 companies

Red stars mark editorial starting points for specific needs. Read the reason on each card. This is a researched shortlist, not a tested effectiveness ranking. Affiliate revenue does not determine inclusion; these links do not currently earn OutOfAI℠ referral commissions.

Entra · Purview · Defender · Agent 36501

★ Microsoft

Employees using AI · AI agents · Testing & oversight

Agent identities and access, sensitive business data, and security monitoring across supported Microsoft services.

Who should evaluate it
IT and security teams already working in a Microsoft environment.
Integration checks
Map the required Entra, Purview, Defender and Agent 365 capabilities to your licenses, supported agents and administrator configuration.
Pricing & access
Licensing varies by product and capability. Confirm the complete entitlement and deployment cost with Microsoft.
Why the red star?

A starting point for existing Microsoft customers: first assess the controls and integrations available in your environment.

Sources reviewed October 22, 2026 · Ordinary provider links

Prisma AIRS02

★ Palo Alto Networks

AI agents · AI apps & models · Testing & oversight

AI discovery, model risk assessment, application testing, runtime safeguards and agent controls.

Who should evaluate it
Enterprise security teams overseeing AI applications, models and agents across their lifecycle.
Integration checks
Confirm supported cloud environments, model formats, deployment points and logging integrations for your workload.
Pricing & access
Request a scoped quote and confirm which Prisma AIRS components and usage are included.
Why the red star?

A starting point for a broad AI security evaluation spanning discovery, testing and runtime protection.

Sources reviewed October 22, 2026 · Ordinary provider links

AI Defense03

Cisco

Employees using AI · AI apps & models · Testing & oversight

AI application discovery, model and application validation, and runtime controls for risks such as prompt injection and data leakage.

Who should evaluate it
Security and platform teams adopting AI services or deploying their own AI applications.
Integration checks
Check supported cloud connections, application integration options and where inspection occurs in your architecture.
Pricing & access
Ask Cisco for current packaging, required integrations and a deployment-specific quote.

Sources reviewed October 22, 2026 · Ordinary provider links

AI Security · Lakera04

Check Point

AI agents · AI apps & models · Testing & oversight

AI applications and agents against prompt injection, data exposure and risky interactions with connected tools.

Who should evaluate it
Application security teams and organizations deploying agents with business-system access.
Integration checks
Review API integration, supported agents and MCP connections, policy enforcement and required data flows. Lakera is part of Check Point.
Pricing & access
Confirm current product packaging and quote or plan terms for the capabilities you need.

Sources reviewed October 22, 2026 · Ordinary provider links

AI and agent security05

★ Noma Security

AI agents · AI apps & models · Testing & oversight

Visibility into AI assets, agents and connected tools, with posture assessment, testing and runtime controls.

Who should evaluate it
Security teams trying to discover and govern agents across a growing AI environment.
Integration checks
Confirm discovery coverage for your models, agents and MCP servers, plus the permissions required by each connector.
Pricing & access
Request a demonstration and a quote based on your assets, integrations and required controls.
Why the red star?

A starting point when your first problem is understanding which AI agents and connected tools exist and what they can access.

Sources reviewed October 22, 2026 · Ordinary provider links

Model scanning · AI runtime security06

HiddenLayer

AI apps & models · Testing & oversight

Model files and deployed AI systems through model scanning, discovery and runtime security capabilities.

Who should evaluate it
Teams building, acquiring or deploying machine-learning models and AI applications.
Integration checks
Check supported model formats, registries, CI/CD pipelines and inference environments before a pilot.
Pricing & access
Request pricing for your model scanning and runtime scope; confirm usage limits and deployment requirements.

Sources reviewed October 22, 2026 · Ordinary provider links

AI Security07

Zscaler

Employees using AI · AI agents · AI apps & models

Employee access to AI services, sensitive-data exposure and AI application interactions through access and runtime controls.

Who should evaluate it
IT and security teams managing employee AI use and enterprise access policies.
Integration checks
Confirm traffic routing, endpoint and identity prerequisites, inspection settings and application coverage.
Pricing & access
Request a quote and confirm the AI security components included in your proposed subscription.

Sources reviewed October 22, 2026 · Ordinary provider links

Inbound email security08

Abnormal AI

Email & transactions

Business email against phishing, impersonation and business email compromise, including attacks that may be assisted by AI.

Who should evaluate it
Security teams protecting employee email and organizations concerned about payment-redirection fraud.
Integration checks
Review supported email platforms, API permissions, remediation settings and alert handling with your mail administrators.
Pricing & access
Request pricing for your email environment and the products included in the proposed deployment.

Sources reviewed October 22, 2026 · Ordinary provider links

02 / Could this happen in your business?

A document should not authorize a payment.

Imagine an accounting agent reading a vendor document. Hidden instructions try to make it change the vendor’s bank details. The document is evidence to review; it is not authority to act.

The exposure

Too much access. Too little review.

An agent that can both read untrusted documents and change payment records creates a path to unauthorized actions. An email security tool alone does not govern every action that agent can take.

The controls

Put decisions behind real checks.

Remove payment-write access from the reading agent. Require a separate human approval, verify bank changes through a known contact channel, and log the request, decision and action. Test that these controls cannot be skipped.

Illustrative scenario, not a report of an incident. No single filter or product eliminates every AI security risk.

03 / Turn a concern into a plan

Your enterprise AI-security checklist.

Download checklist ↓

Use these twelve checks to prepare a pilot with your IT, security and business owners. Assign an owner, evidence and a review date for each item.

Open the 12-point checklist
  1. Inventory your AI

    List approved AI tools, models, agents, owners, business tasks and connected systems. Include employee-installed tools and MCP servers.

  2. Map the data

    Record what data each system reads, sends and retains; confirm approved processing locations and retention settings.

  3. Limit agent access

    Use distinct identities, least-privilege permissions and scoped credentials. Keep secrets out of prompts. Set expiry and revoke unused access.

  4. Approve sensitive actions

    Require an authorized person to approve payments, bank-detail changes, external sharing and other high-impact actions. Enforce approval outside the model.

  5. Treat inputs as untrusted

    Treat email, documents, web pages and tool output as data that may contain malicious instructions. Separate instructions from source content.

  6. Control connected tools

    Allow only reviewed tools and MCP servers. Validate tool inputs and outputs, limit destinations and record changes to permissions.

  7. Check the supply chain

    Review model and software provenance, scan supported model files and dependencies, and patch the components you operate.

  8. Test representative risks

    In an authorized test environment, assess prompt injection, data leakage, excess permissions and unauthorized actions. Record failures and false positives.

  9. Verify payment changes independently

    Use a known telephone number or established trusted channel to confirm changed payment details. Do not rely on a number in the incoming request.

  10. Keep usable records

    Log relevant access, tool calls, approvals and security events. Protect logs, restrict access and apply an appropriate retention policy.

  11. Practice stopping an agent

    Assign an incident owner. Test disabling an agent, revoking credentials, isolating connected systems and preserving evidence.

  12. Review before expanding

    Check current vendor coverage, licensing and data terms. Measure test results, staff effort and costs before scaling the pilot.

Learn the threat patterns

Use the OWASP agentic security guidance to structure questions about agent permissions, tool use and oversight.

OWASP agentic application risks ↗

Follow the companies

Explore original security research and announcements, alongside AI product and infrastructure updates.

Security company newsrooms →