AI security and governance

What you believe about AI in your business, tested.

For companies using AI tools, and for companies shipping AI features. Which tools people actually use, what data goes into them, what your vendors do with it, and whether the AI features you ship can be manipulated through their inputs: we check those the same way we check everything else, with evidence, owners, and dates to look again.

Start the free AI discovery Book a 15-minute call

The threat landscape

AI changes what is true faster than any policy can keep up

Most AI risk in a business is not exotic. It is an assumption that stopped being true: a tool someone adopted, a data flow nobody mapped, a vendor that quietly added a model. The Lens treats each of these as a belief with an owner and tests it.

Belief

"Nobody uses AI tools we have not approved."

Tested by second testimony and sign-in records. In small businesses this belief fails more often than any other.

Belief

"Customer data never goes into public AI tools."

Tested against the approved-tool list, data-handling terms, and a walkthrough of how people actually work.

Belief

"Our vendors have told us where they use AI."

Tested against contracts and vendor disclosures. Usually untested, which is itself the finding.

Belief

"Our AI features cannot be manipulated through their inputs."

For companies that ship AI: tested against the OWASP LLM Top 10 and MITRE ATLAS, with red-team results where they exist.

Technical Prompt injection, model evasion, and supply chain attacks that firewalls and endpoint detection were never designed to catch.
Operational Shadow AI, unmonitored model behavior, and autonomous agents acting without human oversight.
Strategic Without AI-specific threat modeling, leadership decides with incomplete information.
Regulatory The EU AI Act and state-level laws create obligations that need real capability, not just policy documents.

Full-spectrum coverage

Seven frameworks, used as vocabulary, not as the goal

Beliefs are tested first. Frameworks come second, as the vocabulary a regulator, customer, or insurer will use when they ask what you did. Regulation is moving, and we will tell you which apply to you.

Maturity model

Gartner CISO MCP 2026

Measures your AI security capability from reactive to strategic across five stages.

  • Stage 1, Reactive No formal AI security program; ad-hoc responses to incidents
  • Stage 2, Aware AI inventory initiated; basic risk identification underway
  • Stage 3, Proactive Governance policies in place; systematic risk assessment across AI systems
  • Stage 4, Managed Continuous monitoring; AI risk integrated into enterprise risk management
  • Stage 5, Strategic AI security embedded in business strategy; board-level reporting
Risk management

NIST AI RMF

A structured lifecycle for identifying, quantifying, and mitigating AI-specific threats.

  • Govern Establish policies, roles, and accountability structures for AI risk
  • Map Identify and categorize AI systems, data flows, and context of use
  • Measure Assess and quantify risks using metrics, testing, and evaluation
  • Manage Prioritize, respond to, and monitor risks throughout the AI lifecycle
Vulnerability assessment

OWASP LLM Top 10 2025

The ten most critical attack vectors targeting LLM-powered applications.

  • Prompt injection Manipulating LLM behavior through crafted inputs to bypass safety controls
  • Sensitive information disclosure Models leaking training data, PII, or proprietary information
  • Supply chain vulnerabilities Compromised models, datasets, or plugins introducing hidden risks
  • Data & model poisoning Corrupting training data to manipulate model outputs at scale
  • + 6 more Excessive agency, overreliance, insecure output handling, and model theft
Agent security

OWASP Agentic Apps 2026

Attack surface analysis for autonomous AI systems making real-world decisions.

  • Excessive autonomy Agents making high-impact decisions without human approval checkpoints
  • Trust boundary violations Agents accessing systems or data beyond their intended scope
  • Cascading hallucinations Agent errors compounding through multi-step workflows into real actions
  • Tool misuse Agents invoking external tools or APIs in unintended or adversarial ways
Threat intelligence

MITRE ATLAS

15 adversarial tactics and 66 techniques used to attack AI systems in the wild.

  • Reconnaissance Probing AI systems to understand model architecture and training data
  • Model evasion Crafting inputs that cause misclassification while appearing normal
  • Model theft Extracting model weights or behavior through repeated API queries
  • ML supply chain Compromising model registries, training pipelines, or deployment
  • + 11 more tactics Persistence, exfiltration, initial access, and impact
Regulatory compliance

EU AI Act

Risk classification, conformity assessment, and enforcement timeline for AI systems.

  • Unacceptable risk Banned practices: social scoring, real-time biometric surveillance, manipulation
  • High risk Healthcare, employment, finance, and law enforcement require conformity assessment
  • Limited risk Chatbots and deepfakes carry transparency obligations
  • Minimal risk No specific obligations; voluntary codes of conduct encouraged
  • Penalties Up to 7% of global annual turnover or 35M EUR
Operational controls

ISO 42001

A certifiable management system for AI development, deployment, and monitoring.

  • AI policy Formal commitment to responsible AI development and deployment
  • Risk treatment Systematic identification, assessment, and mitigation of AI risks
  • Data governance Controls for data quality, provenance, bias detection, and privacy
  • Continuous improvement Audit cycles, performance evaluation, and management review
  • Certification Third-party auditable proof of AI governance maturity

How it is delivered

Free to start, scoped when it matters

Free

AI Security Discovery

A 15-question diagnostic that maps where your AI beliefs are most likely to be untested, across the seven frameworks. Anonymous, no account, immediate results.

About 7 minutes
Start the free discovery
With a Security Health Check

AI questions, included

Every Security Health Check includes a few questions about AI use. A deeper review adds an inventory of approved tools, a look at what your vendors do with AI, and a walkthrough of where your data goes.

For businesses with up to 100 people
See the Health Check
With Strategic Security Advisory, or on its own

AI governance review

A 50-question assessment across inventory, governance, risk management, technical controls, and compliance, interviews with the people who build or buy AI, and red-team results where they exist. Mapped to all seven frameworks.

For growing companies and enterprises
See our services

Start with what you believe

The free 15-question discovery takes about seven minutes, needs no account, and tells you which AI beliefs are most likely untested. From there, the module fits inside whichever engagement fits your company.

Take the free AI discovery Book a 15-minute call