LLM and vendor risk
Who owns each vendor, where data is processed, whether your inputs train their models, and what their security posture actually is.
Defend
You deployed AI. Who's securing it?
A fixed-scope project that answers three questions: which AI tools and agents your company actually uses, what data and actions those agents can reach, and which of those paths put you at risk. In two to four weeks you get a full inventory, risk-ranked findings, an acceptable use policy, and a 90-day remediation plan that includes agent and MCP controls.
Shadow AI discovery
The AI you approved is not the AI in use.
Every company has two AI footprints: the one IT approved, and the bigger one employees and agents built themselves.
~55%of employees admit to using AI tools their company never approved.UHY middle-market research, 2026
~35%of the data employees paste into AI tools is sensitive: client records, financials, source code, credentials.Cyberhaven data flow analysis
Discovery pairs log analysis with staff interviews: personal ChatGPT accounts doing client work, free tools with training rights over your data, AI features switched on inside software you already own, and agents with tools (MCP and similar connectors) that can take real actions with incomplete visibility. Excessive agency is when those agents are authorized to do more than policy intended. For a preview of the employee side, read our guide to shadow AI and what to do about it.
What we assess
Six surfaces, tested against how attackers and auditors think.
Data flows
We map what leaves your boundary and where it lands: prompts, uploads, connected drives, integrations, and agent tool calls, ranked by sensitivity.
Prompt injection and tool paths
If chatbots, copilots, or agents touch your data, we test how they handle hostile input designed to hijack them, and we exercise the risky tool paths those models can call.
MCP blind spots
We inventory agent workflows, tool actions, and permissions. MCP and similar connectors can act on mail, files, tickets, and production systems while the chat UI shows only a slice of what the agent can do.
Excessive agency
Who can connect an agent to company systems, and with what rights. Excessive agency is an agent authorized to do more than policy intended. Overshared drives and wide tool grants turn one prompt into a company-wide action.
Acceptable use policy
We review or write your policy: approved tools, agent permissions, prohibited data and actions, and consequences, in language staff will read.
Governance is mapped against NIST AI RMF, the standard insurers, auditors, and enterprise customers increasingly ask about. Agent and MCP controls sit in that same picture, not in a side memo.
Deliverables
What you hold at the end.
AI inventory and exposure map
Every tool, account, agent workflow, and MCP or similar tool connection in use, with a map of what sensitive data goes where and which actions those agents can take. This becomes your ongoing register.
Risk-ranked findings report
Each finding scored by likelihood and impact, in plain language for executives with technical detail attached. MCP blind spots and excessive agency sit in the same ranking as vendor and data-flow risk. No padding, no scare graphics.
Policy and 90-day plan
An acceptable use policy ready to adopt, plus a remediation plan with owners and sequence that includes agent and MCP controls, presented in a working session.
The other side of AI risk
Attackers adopted AI faster than you did.
Attackers use the same models your employees experiment with to write flawless phishing, clone executive voices, and stage deepfake video calls that approve wire transfers.
$3.04Bin business email compromise losses logged by the FBI in 2025 alone.FBI IC3 annual report, 2025
We test how your payment approvals, identity verification, and executive communications hold up against AI-assisted fraud, and we run deepfake tabletop exercises through our cybersecurity consulting practice.
Questions
Straight answers on AI security assessments
How long does an AI security assessment take?
Two to four weeks for most companies with 25 to 1,000 employees. Week one is discovery, weeks two and three are analysis and testing of data flows and risky tool paths, and you get the written report and remediation plan in a working session, not an inbox drop.
We only use ChatGPT and Copilot. Do we still need this?
That is the exact profile that needs it. The risk is rarely the sanctioned chat tool; it is personal accounts, free tools with training rights over your data, AI features switched on inside software you already own, and agents with tool access that can act without a complete picture of what they can reach.
What do we get at the end?
Five deliverables: an inventory of every AI tool, agent workflow, and MCP or similar tool connection in use, a data exposure map, a risk-ranked findings report, an acceptable use policy ready to adopt, and a 90-day remediation plan with owners that includes agent and MCP controls.
What are MCP blind spots and excessive agency?
MCP and similar connectors let agents call tools that take real actions: mail, tickets, files, browsers, and production systems. A blind spot is a tool path or permission you cannot see from the chat UI. Excessive agency is an agent authorized to do more than policy intended. We inventory those workflows, test the risky paths, and write the limits into policy and the 90-day plan.
Will you tell us to ban AI tools?
No. Bans push AI use onto personal devices where you have no visibility. We pair controls with a sanctioned path: approved tools and agent permissions, clear rules for what data and actions are allowed, and training so people use them well.
Book a Discovery Call
Thirty minutes with a senior practitioner. Map the agents and tools you already run. No obligation.
Book a Discovery Call