AI Security Assessment: 12 Best Services and Firms Ranked (2026)
These are the 12 best AI security assessment services for 2026, ranked by how well each firm can find and prioritize the real risks in the AI you use and build. An AI security assessment is a scoped review of your AI estate: which models, copilots, agents and data flows you have, how an attacker could abuse them, how well your controls hold up and what to fix first. Firms differ in where they look hardest, from AI threat modeling and LLM application testing to AI governance gap assessments and audits that feed a certificate. The ranking shows what each AI security assessment company actually delivers, then covers scope, frameworks, process, pricing and how to pick one.
An AI security assessment answers three questions: what AI do we have, how could it be attacked or misused, and are our controls good enough. A good one starts with discovery (including shadow AI and AI features inside SaaS tools), then reviews architecture, data access, identities and configurations, tests the riskiest systems and ends with a prioritized roadmap. It is broader than a pen test and narrower than a full governance program.
There are four common types. AI threat modeling maps how attackers would go after a specific application or agent. LLM application testing attacks a live system for prompt injection, data leakage and unsafe tool use. AI governance gap assessments compare policies and controls with frameworks such as the NIST AI RMF, ISO/IEC 42001 and the EU AI Act. Third-party or AI vendor security assessments check the AI suppliers you buy from. Most firms below run all four, but each has a center of gravity.
AI red teaming and AI audits that lead to ISO 42001 certification or HITRUST reports
SaaS and healthcare tech companies whose customers want a certificate or attestation, not just a test report
Project and audit fees; quote only
Tampa, Florida
7.6
Reviews
The top AI security assessment companies, reviewed
Each review covers what the company sells, who it suits, how it prices and where it falls short.
Rank #1
1. Palo Alto Networks Unit 42
Threat-informed AI security assessment across AI use, development, runtime and governance
9.3
Unit 42's AI Security Assessment is the most complete packaged offer we found. It reviews employee AI use, AI development infrastructure and data, runtime threats to deployed AI apps through threat modeling, governance guardrails and how AI can help the SOC. The output is a threat-informed report and a risk-prioritized roadmap.
Best for
Large enterprises that want one assessment covering employee AI use and AI they build
Headquarters
Santa Clara, California
Founded
2014
Ownership
Part of Palo Alto Networks (NASDAQ: PANW)
Key products
Unit 42 AI Security Assessment, Unit 42 Frontier AI Defense, Continuous Frontier AI Defense, Unit 42 Retainer
Pricing model
Project fee or prepaid Unit 42 Retainer hours; quote only
The engagement runs in four phases, which Unit 42 calls plan, explore, evaluate and activate: identify risks tied to your AI plans, run workshops on business impact, test the current design against best practice and define the target posture. Its scope, per the service page, includes AI applications, models, data flows, identities, third-party services and governance controls, with a focus on data leakage, insecure integrations and excessive access. Threat intelligence from Unit 42's incident response work shapes which attacks it prioritizes.
Unit 42 added Frontier AI Defense in April 2026 and an annual subscription, Continuous Frontier AI Defense, in September 2026, for ongoing AI-driven exposure testing after the first assessment. Recommendations naturally lean toward Prisma AIRS and Cortex, so ask for product-neutral findings if you use other tools.
AI threat modeling, AI/ML secure development reviews and AI red teaming
9.1
NCC Group brings AI and ML specialists into a large penetration testing and threat modeling practice. It reviews AI development pipelines, runs AI/ML red teaming against OWASP LLM and ML Top Ten risks and offers AI readiness assessments. After selling its Escode escrow business in May 2026 it operates purely as a cyber firm.
Best for
Companies building their own AI products that need design-stage threat models and testing
Headquarters
Manchester, United Kingdom
Founded
1999
Ownership
Public (LSE: NCC)
Key products
AI security assessments, AI/ML red teaming, AI/ML secure development lifecycle review, AI readiness, threat modeling
Pricing model
Project fees and retainers; quote only
Where NCC Group stands out is early in the lifecycle. Its threat modeling and secure development reviews look at how models are trained, versioned, deployed and connected to tools, which catches design flaws that a late pen test would only find as symptoms. Its research team publishes on AI attacks, so testers bring current techniques.
NCC Group is strongest for teams that build AI. Companies that mainly consume third-party copilots and SaaS AI may get more from a discovery-first assessment.
AI security assessments built on Google's Secure AI Framework, plus AI red teaming
9.0
Mandiant's AI security consulting portfolio assesses AI programs against Google's Secure AI Framework (SAIF), red teams generative AI in production, reviews AI applications and prepares SOC teams for AI incidents. It suits companies that want the assessment shaped by attacks seen in real breaches.
Best for
Enterprises that want front-line attacker knowledge and a SAIF-based assessment
Headquarters
Reston, Virginia
Founded
2004
Ownership
Subsidiary of Google (Alphabet, NASDAQ: GOOGL) since 2022
Key products
AI Security Assessment (SAIF), Red Teaming for AI, AI application review, SOC readiness and AI incident response playbooks
Pricing model
Project fees or Mandiant retainer hours; quote only
The SAIF assessment looks at controls across the AI lifecycle, from data and model supply chain to deployment and monitoring, and the red team then tests whether those controls stop realistic attacks. Mandiant's incident response playbooks for AI are written to work across platforms, which matters for companies that run models from several providers.
Being part of Google means some buyers worry about bias toward Google Cloud. In practice the SAIF method is cloud-neutral, but ask that recommendations name controls rather than products.
Telemetry-led discovery of AI use, then expert review of AI architecture and controls
8.9
CrowdStrike's AI Systems Security Assessment starts by finding where AI runs across cloud, SaaS and internal environments, including LLMs, copilots, agents and unsanctioned tools, using Falcon platform telemetry. Experts then review architecture, data access, configurations and user controls with attacker tactics in mind, and deliver a report plus stakeholder workshops on guardrails.
Best for
CrowdStrike customers who need to find all their AI before they can secure it
Headquarters
Austin, Texas
Founded
2011
Ownership
Public (NASDAQ: CRWD)
Key products
AI Systems Security Assessment, AI Red Team Services, Shadow AI Visibility Service, Frontier AI Readiness and Resilience Service
Pricing model
Services engagement; quote only
Discovery is the hard part of most AI assessments, and CrowdStrike's telemetry gives it a head start when Falcon is already deployed. The assessment names risks such as prompt injection, model misuse and insecure configuration, and CrowdStrike's AI Red Team Services, launched in November 2024, can follow with hands-on testing of the riskiest systems.
The public service page does not name the frameworks behind its findings, so ask how results map to the OWASP LLM Top 10, MITRE ATLAS or the NIST AI RMF. Value is highest for existing Falcon customers.
Human-led AI/ML security assessments and LLM application pentesting
8.7
Bishop Fox runs AI/ML security assessments with senior human testers. Its work covers LLM endpoints for jailbreaks, context leaks and secrets extraction, cloud AI risks such as denial of wallet, GenAI architecture reviews and AI-focused red teams that move from phishing to model artifacts.
Best for
Companies whose AI features handle sensitive data or take actions on users' behalf
The architecture review is a useful first step for teams early in a GenAI build. It looks at where prompts, retrieval data, tools and secrets sit and where trust boundaries break, before testers go after the live system. Cosmos, its continuous testing service, can keep watch on internet-facing AI endpoints afterward.
Bishop Fox is an offensive security firm first. For governance gap work against ISO 42001 or the EU AI Act, pair it with a governance firm.
GenAI pentesting, attack path mapping and AI threat modeling across the full AI stack
8.6
Praetorian tests the whole AI stack: interfaces, APIs, RAG sources, agents, plugins and vector databases. Its engagements include GenAI penetration tests guided by the OWASP Top 10 for LLM Applications, attack path mapping, red team operations, AI-specific threat models and gap analysis against the NIST AI RMF.
Best for
Engineering-led companies that want LLM app testing tied to threat models and fix verification
Headquarters
Austin, Texas
Founded
2010
Ownership
Private
Key products
GenAI Penetration Test, GenAI Attack Path Mapping, GenAI Red Team Operation, AI threat modeling, Augustus LLM scanner, Praetorian Guard
Pricing model
Project fees; Praetorian Guard by annual subscription; quote only
Praetorian's method maps adversary emulation to MITRE ATLAS and references Google's Secure AI Framework, and its deliverables include business-risk scoring, threat models and follow-up verification of fixes. It built Augustus, an open-source LLM vulnerability scanner that runs more than 210 adversarial attacks against 28 LLM providers, and includes it in the Praetorian Guard platform.
Its GRC and gap analysis work is lighter than its offensive work. Companies that need a board-level governance program should pair it with a governance firm.
Research-grade AI/ML security assessments from the model up
8.5
Trail of Bits assesses AI and ML systems at a depth few firms match, including training pipelines, model file formats and ML supply chains. It has worked with large AI platforms, including a published engagement with Hugging Face.
Best for
AI labs, model hosts and ML platform teams that need code-level review
Headquarters
New York, New York
Founded
2012
Ownership
Private
Key products
AI/ML security assessments, ML pipeline and supply chain review, AI red teaming, open-source AI security tools
Pricing model
Consulting engagement; quote only
Where most assessments stop at the application, Trail of Bits reviews the code and infrastructure that produce and serve models. That catches risks like unsafe model serialization, dependency compromise and flaws in inference servers, which matter most to companies that train, host or distribute models.
Its depth suits technical buyers. Companies that mainly consume AI through SaaS or copilots will not use most of what it offers, and it does not run governance programs.
AI assessments across 14 engagement types, from agents and models to AI in devices
8.3
IOActive sorts its AI work into 14 engagements across four stages: before you build, while you build, before you ship and after you ship. It tests seven layers of the AI attack surface, from user input and prompts to RAG, agents and tools and infrastructure, and maps work to the NIST AI RMF, MITRE ATLAS, the EU AI Act and ISO/IEC 42001.
Best for
Manufacturers and product companies putting AI into devices, vehicles or embedded systems
Headquarters
Seattle, Washington
Founded
1998
Ownership
Private
Key products
AI Agent Security Assessment, AI Red Teaming, AI Application Security, AI Model Assessment, AI Data and Privacy, Continuous AI Assessment
Pricing model
Project fees and continuous assessment; quote only
IOActive is a research-led firm with hardware and embedded labs in Seattle, Cheltenham and Madrid, which sets it apart when AI runs on a device rather than in the cloud. Its engagements include model custody and infrastructure reviews, retesting with replayable findings and a regression harness, so fixes can be checked over time.
Its public material on AI is newer than its long hardware record, and it publishes fewer AI case studies than the firms above. Ask for examples close to your system.
AI threat and risk assessments, AI application testing and AI strategy reviews
8.1
Kudelski Security's AI Threat and Risk Assessment runs in five phases: inventory AI systems (including third-party and shadow AI), map use cases and dependencies, assess risks by category, rate and prioritize them, then report. Its AI red team runs separate adversarial testing of LLM applications.
Best for
Mid-sized and large companies that want a structured AI risk register plus LLM app testing
Headquarters
Cheseaux-sur-Lausanne, Switzerland (US base in Phoenix, Arizona)
Founded
2012
Ownership
Division of Kudelski Group (SIX: KUD)
Key products
AI Threat and Risk Assessment, AI Application Security Testing, AI Strategy and Security
Pricing model
Project fees; quote only
The output is a standardized assessment report, a prioritized roadmap, a governance framework and policy templates, with workshops to align IT, security, data science and compliance. Kudelski references OWASP, the NIST AI RMF, MITRE ATT&CK, ISO 42001 and the EU AI Act. That mix suits buyers who want one engagement to cover both risk register and governance starting points.
Kudelski is better known for managed detection and response than for AI testing, so check the AI red team's size and experience. In November 2026 it joined CrowdStrike's Project QuiltWorks.
AI/ML penetration testing at scale with consultants and a vetted researcher crowd
8.0
NetSPI tests AI and LLM applications and machine learning models through its PTaaS platform. Its merger with Synack, closed in October 2026, adds a vetted researcher crowd and Sara, an AI pentesting agent whose findings are validated by human experts.
Best for
Companies that need many AI apps tested on a schedule through one platform
Headquarters
Minneapolis, Minnesota
Founded
2001
Ownership
Private (majority owned by KKR); merged with Synack in October 2026
Key products
AI/ML penetration testing, NetSPI PTaaS, Synack Sara AI pentesting, Synack Platform
Pricing model
Per test or subscription; quote only
NetSPI's strength is volume and cadence. Findings land in one platform with retesting, which suits companies shipping AI features every few weeks. Its AI/ML testing covers machine learning models and LLM implementations alongside the usual web and API checks.
The merger is recent, so ask how the two platforms and contracts will combine. NetSPI is a testing firm, not a governance advisor.
AI security testing and AI risk assessments mapped to compliance frameworks
7.8
Coalfire combines AI security testing (red teaming, adversarial testing and LLM tests for prompt injection, misuse and hallucination) with control testing mapped to the NIST AI RMF, HITRUST's AI risk management framework, PCI DSS and SOC 2. It also holds ANAB accreditation for ISO 42001 and provisional AIUC-1 accreditation.
Best for
Regulated companies that need AI test results mapped to PCI DSS, HITRUST, SOC 2 or ISO 42001
Headquarters
Westminster, Colorado
Founded
2001
Ownership
Private (Apax Partners funds since 2020)
Key products
AI security testing (red teaming, adversarial and LLM testing), AI model testing, AI risk management advisory, AIMS+, AIUC-1 assessments
Pricing model
Project fees; quote only
For payment, healthcare and federal buyers, Coalfire's value is translation: it turns AI test findings into control evidence an auditor will accept. It has a specific service for AI that touches payment data, covering fraud models and customer service bots.
AI testing is one part of a broad compliance practice, so check the testers' AI experience. And because Coalfire also certifies, keep its assessment and certification roles separate on the same scope.
AI red teaming and AI audits that lead to ISO 42001 certification or HITRUST reports
7.6
Schellman is an audit firm with an AI practice. It runs AI red teaming of generative AI systems for prompt injection and data leakage, was the first ANAB-accredited ISO 42001 certification body, and is a HITRUST CSF assessor, which matters for buyers looking at the HITRUST AI security assessment.
Best for
SaaS and healthcare tech companies whose customers want a certificate or attestation, not just a test report
Headquarters
Tampa, Florida
Founded
2002
Ownership
Private (Goldman Sachs Alternatives strategic investment announced March 2026)
Key products
AI Red Teaming, ISO/IEC 42001 certification, HITRUST assessments, AI governance services
Pricing model
Project and audit fees; quote only
The HITRUST AI Security Assessment, announced in November 2024, adds up to 44 AI-specific requirements on top of an e1, i1 or r2 assessment and must be performed by an authorized external assessor. Schellman's guidance explains how those requirements overlap with ISO 42001, which helps healthcare technology firms plan one evidence set for both.
Its AI red teaming is narrower than the offensive specialists'. And Schellman cannot both advise on and certify the same system, so use it for testing or for certification on a given scope, not both. Confirm its HITRUST authorization covers the AI assessment before you sign.
Decide which question you need answered Do you need to know what AI you have, whether a specific app can be broken, or whether your controls meet a framework? Those are discovery, testing and gap assessment, and firms are strongest at one of them. Write the question down before you call anyone.
Insist on discovery first Inventories built from interviews alone tend to miss AI, especially AI features inside SaaS tools and agents built by business teams. Pick a firm that starts with discovery, through telemetry, SaaS logs, procurement records or interviews, or run a shadow AI tool before the assessment.
Ask which frameworks findings map to Good AI security assessment services map findings to the OWASP Top 10 for LLM Applications, MITRE ATLAS and the NIST AI RMF, and to ISO/IEC 42001 or HITRUST if you need certification later. Unmapped findings are harder to track and to explain to auditors.
Check who does the testing Ask for the names and backgrounds of the testers. AI assessments need people who understand both application security and how models, retrieval and tool calls behave. Ask how much of the testing is automated and how findings are validated.
Agree deliverables and retest terms A useful report has an AI inventory, threat models for key systems, findings ranked by business impact, a roadmap and a free or discounted retest. Agree these in the statement of work.
Watch for product tie-ins Several leading AI security assessment companies also sell AI security platforms. That is fine if recommendations name controls first and products second. Ask for a product-neutral roadmap in writing.
Pricing
How AI security assessment pricing works
None of the firms on this list publishes a price for an AI security assessment. Most quote a fixed project fee based on how many AI systems are in scope, how many get hands-on testing, whether discovery is included and how many workshops you want. Enterprise buyers can often pay from prepaid incident response retainer hours, which Unit 42 and Mandiant both allow for proactive services.
Ongoing options are growing. Unit 42 sells Continuous Frontier AI Defense on an annual subscription, NetSPI sells testing through its PTaaS platform, and Praetorian bundles continuous testing into Praetorian Guard. HITRUST AI security assessments and ISO 42001 audits are priced like other audits, by assessor time and scope. Ask every firm for a day count, the number of systems tested and the retest policy so quotes can be compared.
Process
How an AI security assessment works, step by step
Most engagements follow the same arc, even if each firm names the phases differently. Unit 42 calls them plan, explore, evaluate and activate. Kudelski Security uses inventory, map, assess, rate and report. The work underneath is similar.
A typical scope for a mid-sized company covers the AI inventory, three to five systems for deep review and one or two for hands-on testing. Larger programs repeat the cycle per business unit.
Scoping and discovery List AI apps, copilots, agents, models, data sources and AI suppliers, including shadow AI found through telemetry, SaaS logs and interviews.
Context and threat modeling Map each priority system's data flows, identities, tools and trust boundaries, and the attacks most likely against it given your industry and region.
Control review Check configurations, access, logging, guardrails, data handling and governance controls against the chosen frameworks.
Hands-on testing Attack the riskiest systems for prompt injection, data leakage, insecure tool use, excessive agent permissions and model supply chain issues.
Risk rating Score each finding by likelihood and business impact, taking existing controls into account.
Report and roadmap Deliver findings, threat models and a prioritized roadmap, then retest fixes.
Frameworks
AI security assessment frameworks firms use
There is no single AI security assessment framework. Firms combine several, and you should know which ones before you sign. The OWASP Top 10 for LLM Applications is the usual checklist for application testing, covering prompt injection, sensitive information disclosure, supply chain issues, excessive agency and similar risks. MITRE ATLAS catalogs real attacker tactics and techniques against AI systems and is used for adversary emulation and threat models.
For control and governance reviews, the NIST AI Risk Management Framework is the most common reference in the US, and ISO/IEC 42001 is the certifiable management system standard. Google's Secure AI Framework (SAIF) shapes Mandiant's assessments and appears in Praetorian's method. European buyers add the EU AI Act, and healthcare buyers often add HITRUST's AI security requirements.
RAND publishes a free AI Security Guide and Risk Assessment Tool that recommends controls based on a questionnaire about your AI system. It is a good way to prepare for an AI security assessment, and RAND notes it does not replace expert review.
HITRUST
HITRUST AI security assessment and certification
HITRUST announced its AI Security Assessment in November 2024. It adds up to 44 AI-specific security requirements to an existing HITRUST e1, i1 or r2 assessment, with the exact number depending on the type of model and data in use. Built on an e1 or i1 it is called ai1, and built on an r2 it is called ai2. Certification requires passing both the base assessment and the AI requirements.
The assessment must be performed by a HITRUST authorized external assessor organization, which makes it different from a consulting-led AI security assessment. Healthcare technology companies often do both: a technical assessment from a testing firm to find and fix issues, then the HITRUST AI assessment to give customers a formal report. Schellman's analysis notes that some of the HITRUST AI requirements map to ISO 42001, so one evidence set can serve both.
Third parties
AI vendor security assessments: checking the AI you buy
Much of the AI risk in a typical company sits with suppliers: SaaS tools that added AI features, model APIs and agent platforms. An AI vendor security assessment asks how a supplier trains on customer data, where prompts and outputs are stored, which sub-processors and model providers sit behind the product, how it defends against prompt injection and whether it holds ISO 42001, SOC 2 or HITRUST AI evidence.
Most companies run these through their existing third-party risk program with an AI addendum to the security questionnaire. AI vendor security assessment tools, such as questionnaire automation platforms, help suppliers answer faster but do not judge the answers. For high-impact suppliers, ask for a summary of the supplier's own AI security assessment or red team report, and test the integration on your side, since many issues come from how a model is connected rather than from the model itself.
If you are on the supplier side, an independent assessment from one of the firms above, or an ISO 42001 certificate, shortens these reviews. Our list of ISO 42001 certification bodies covers the accredited auditors.
After the report
From findings to fixes: what to do after the assessment
An assessment is only useful if the roadmap gets funded. Most roadmaps land in three buckets: governance work (policies, an AI inventory owner, review gates), technical controls (guardrails, identity and permission limits for agents, logging) and ongoing testing.
For technical controls, many companies buy a platform. AI security platforms cover discovery, model scanning and runtime protection, AI-SPM tools track misconfigured AI services in the cloud, and shadow AI tools control employee use of unsanctioned AI. Our rankings of AI security platforms, AI-SPM tools and shadow AI tools compare the options. For a short list matched to your assessment findings, you can get a free shortlist from us.
Retest within a quarter. Assessments that are never repeated tend to drift out of date as new agents and features ship, which is why continuous testing subscriptions are growing.
Measuring
Metrics and accuracy: judging an AI-driven security assessment
More firms now use AI agents to run parts of the assessment, from Synack's Sara at NetSPI to Unit 42's frontier AI testing and Praetorian's Augustus scanner. AI-driven security assessment accuracy is a fair concern. Ask what share of findings were validated by a human, how false positives are handled and whether you can see the raw test cases.
Useful metrics for the engagement itself include the share of AI systems in the inventory that were reviewed, the number of high and critical findings per tested system, time to remediate them and the result of the retest. Track the same numbers across assessments to see whether your AI security is improving. For firms that specialize in hands-on testing, see our ranking of AI penetration testing services.
Mistakes
Common mistakes in AI security assessments
Testing only the chatbot. The visible chat interface is rarely the riskiest system. Agents with write access, RAG pipelines fed by shared drives and AI features inside SaaS tools often carry more risk.
Skipping identity and permissions. Many AI incidents come from agents and integrations that hold more access than they need. Make sure the assessment reviews service accounts, API keys and tool permissions, not only prompts.
Treating it as a one-off. Models, prompts and tools change weekly. Plan for a retest and for testing new AI systems before launch.
Mixing assessment and certification. If you will need ISO 42001 or a HITRUST AI report, keep the firm that helps you fix issues separate from the one that certifies you, or the certificate may not be accepted. Our AI security hub explains how assessments fit with platforms and ongoing AI security work.
FAQ
AI security assessment services: FAQ
Answers to the questions buyers ask most about ai security assessment services.
An AI security assessment is a scoped review of the AI systems an organization uses and builds. It inventories them, models likely attacks, reviews controls and tests the riskiest systems, then delivers prioritized findings and a roadmap.
Which companies offer AI security assessment services?
In our 2026 ranking, Palo Alto Networks Unit 42, NCC Group, Mandiant, CrowdStrike and Bishop Fox lead, followed by Praetorian, Trail of Bits, IOActive, Kudelski Security, NetSPI, Coalfire and Schellman. Each has a different focus, from discovery to code-level ML review to audit.
How much does an AI security assessment cost?
Firms do not publish prices. Most charge a fixed project fee set by the number of AI systems in scope, how many get hands-on testing and whether discovery and workshops are included. Enterprises can often use prepaid incident response retainer hours.
How long does an AI security assessment take?
Focused assessments of one or two AI applications often take a few weeks. Enterprise-wide assessments with discovery, workshops and testing of several systems usually take one to three months, depending on scope and access.
What is the difference between an AI security assessment and AI red teaming?
AI red teaming attacks specific systems to see what breaks. An AI security assessment is broader: it includes discovery, architecture and control review, and governance gaps, and may include red teaming as one step.
What is a HITRUST AI security assessment?
It is an add-on to a HITRUST e1, i1 or r2 assessment with up to 44 AI-specific security requirements, announced in November 2024. It must be performed by a HITRUST authorized external assessor and can lead to an ai1 or ai2 certification.
What frameworks are used in an AI security assessment?
Common ones are the OWASP Top 10 for LLM Applications, MITRE ATLAS, the NIST AI RMF, ISO/IEC 42001, Google's Secure AI Framework, the EU AI Act and HITRUST's AI requirements. Ask each firm which ones its findings map to.
Who are the leaders in AI vendor security assessment?
For assessing AI suppliers, most companies extend their third-party risk program with an AI questionnaire, then ask high-impact suppliers for independent evidence such as an ISO 42001 certificate, a HITRUST AI report or a summary from a firm like those ranked here.
Are there AI security assessment tools I can use myself?
Yes. RAND's free AI Security Guide and Risk Assessment Tool recommends controls from a questionnaire, and open-source scanners such as Praetorian's Augustus test LLM endpoints. They help with preparation but do not replace an expert review.
How often should we run an AI security assessment?
At least once a year for the full estate, plus testing of each new high-risk AI system before launch and a retest after major fixes. Fast-moving teams increasingly add continuous testing subscriptions.
Related
More AI security rankings
Related lists and our overall ranking of AI cybersecurity companies.