• 9 min read

30 AI Agent Security Questions for Technical Due Diligence

tl;dr

Identity controls cannot govern agents you do not know about. Runtime discovery and tool-call inspection matter more than per-agent pricing because 82% of enterprises run unknown AI agents.

Featured image for "30 AI Agent Security Questions for Technical Due Diligence"

A Cloud Security Alliance survey found that 82% of enterprises have unknown AI agents operating in their environments, yet most procurement checklists still treat agent security as an identity problem. That mismatch is expensive: Gartner’s 2024 AI Security Report found that 73% of organizations running AI agents experienced at least one security incident related to unauthorized agent actions within their first year of deployment. The questions below close the gap between what vendors claim and what you can verify.

Why Most AI Agent Due Diligence Fails

The standard vendor questionnaire asks who may open a record. An agent review has to ask four things the questionnaire never did: what did the AI see, where could it go, what did it do, and what does an attacker get. OWASP’s Top 10 for Agentic Applications, published December 9, 2025, lists Agent Goal Hijack, Tool Misuse, and Identity & Privilege Abuse as the top three risks for agentic systems — yet most governance purchases still center on identity features and per-agent list price. Cisco’s AI Readiness Index found that only 24% of organizations can control agent actions with proper guardrails and live monitoring, a figure Cisco cites in its announcement of the Astrix deal source.

The market is acquiring identity and non-human identity vendors as the strategic agent security layer. Five independent agent and non-human identity vendors were acquired within 12 months: Veza by ServiceNow (March 2, 2026), Astrix by Cisco (June 29, 2026), Entro by SailPoint (June 29, 2026), Oasis by Cyera (September 3, 2026), and CyberArk by Palo Alto Networks (February 2026); Astrix ended standalone license sales effective June 30, 2026. But OWASP ranks identity below goal hijacking and tool misuse. Runtime tools like Zenity, Noma, and Prisma AIRS watch what agents actually do rather than who they are.

Microsoft Research demonstrated that weight-space backdoors can reprogram fewer than 10 dormant neurons to create sleeper agents that pass all evaluations, with dataset poisoning reaching approximately 78% attack success and model poisoning reaching 84.1% to 94.8% on unseen repositories. Most governance controls and procurement checklists still center on model allow-lists, provider approvals, and eval-based assurance rather than independent runtime observability outside the model. The EU AI Act carries fines reaching €15 million or 3% of global turnover for general-purpose AI violations, and the European Commission has opened enforcement actions questioning OpenAI and Anthropic over AI agent security incidents.

What Questions Should You Ask About Agent Inventory and Scope?

Start with visibility. An undocumented tool, data source, identity, or integration can bypass controls applied to the main agent. The first category of due diligence questions covers scope and inventory: Can you produce a live inventory of every agent, model, prompt configuration, tool, API, database, data source, vendor, integration, dependency, and reachable system? What evidence shows the inventory updates automatically when a team spins up a new agent in a no-code builder or a developer adds an MCP server to their laptop?

Strong answers include a live registry with automated discovery across SaaS builders, local harnesses, and cloud runtimes. Weak answers point to a static spreadsheet or a CMDB that only captures centrally provisioned agents. Ask for a demonstration: show me an agent that was created yesterday in a business unit’s Copilot Studio tenant and appeared in your inventory without manual entry. If the vendor cannot demonstrate automatic discovery of shadow agents, the inventory is incomplete by definition — and per-agent pricing becomes a trap. AgentShield publishes transparent AI agent security pricing at $49 per month for a developer, $199 per month for a team, $749 per month for scale source, and $1,498 per month for Enterprise, but a low per-agent price looks cheap until unknown agents are counted.

How Do You Evaluate Identity and Delegation Controls?

The second category covers identity and delegation. Every agent needs a unique, scoped, non-human identity — but identity alone doesn’t govern what the agent does at runtime. Ask: How are agent identities provisioned, rotated, and revoked? Can you show a configuration export where each agent’s identity is tied to a specific tool scope with per-resource permissions? What happens when an agent spawns a sub-agent — does the sub-agent inherit the parent’s identity, receive a new scoped identity, or run without one?

Microsoft Agent 365 costs $15 per user per month on an annual commitment, or $99 per user per month when included in Microsoft 365 E7, and provides Entra Agent ID for identity within the Microsoft estate source. But identity consolidation via M&A is solving authentication when the dominant incidents are goal hijacking and tool misuse at execution time. A 50-developer team deploying Microsoft Agent 365 at $15 per user per month on an annual commitment spends $9,000 per year on subscriptions alone [50 users × $15/user/month × 12 months]. That covers identity — it doesn’t cover runtime tool-call inspection.

Standard Compute’s Enterprise governance tier costs $1,499 per month and bundles provider allow-lists, admin audit log, request audit export, and SAML SSO; individual governance features are priced from $199 to $399 per month source. Their model allow-list per key costs nothing extra, but the provider allow-list — which controls which providers may see your data — is $199/month. The distinction matters: identity controls who the agent is; provider and model allow-lists control where your data goes.

ToolIdentity & InventoryRuntime Tool-Call ControlPublished Pricing
Microsoft Agent 365Entra Agent ID, Defender inventoryDefender blocks unsafe Foundry/Copilot Studio actions (requires Agent 365 license since July 1, 2026)$15/user/mo annual (andrew.ooo)
AgentShieldAgent identity managementInspection lane blocks high-risk actions at runtime$49–$1,498/mo tiered (agentshield.ai)
Standard ComputePer-key identity with model/provider allow-listsGateway enforces budgets, allow-lists, region rules per call$199–$1,499/mo modular (standardcompute.com)

What Runtime Controls Actually Matter at Execution Time?

The third and fourth categories — runtime controls and containment — are where most evaluations fall short. OWASP’s AI Security Verification Standard (AISVS) 1.0, released in June 2026, provides 191 testable requirements across 12 chapters and is described as suitable for AI agent security procurement. Use its versioned requirements to build a risk-based checklist. Ask: Does the tool inspect every tool call and parameter before execution? Can it block a call based on parameter values, not just tool name? How does it handle prompt injection embedded in untrusted documents, images, or tool responses?

Visual prompt injection in multimodal agents exposes a tool-call control gap: researchers presented Repeat-After-Me: Black-Box Adaptive Visual Prompt Injection at an ACM venue, showing agents caused to expose personally identifiable information and make unauthorized tool calls such as sending messages or invoking external services. Standard input filtering and guardrails operate on text; this attack embeds instructions in image content, bypassing controls most organizations have in place. Unauthorized tool calls represent consequential agent actions — organizations without human approval gates for such actions have no layer of defense when a manipulation succeeds.

Zero-click vulnerabilities in Salesforce Agentforce exposed wider AI agent risk: Zenity Labs disclosed vulnerabilities allowing attackers to silently exfiltrate sensitive CRM data without any interaction from the victim. The attack chain combined prompt injection via Web-to-Lead forms, agent trusting record content as instructions, and the agent’s underlying access to sensitive tool and data permissions. The attacker could perform a successful compromise without direct access to the target organization — the lead submission alone seeded the payload, and normal agent operation did the rest.

Ask for a live demonstration: show me the tool refuse a malicious parameter value in a test tenant. Show me the approval gate trigger on an irreversible action. Show me the kill switch revoke an agent’s credentials in under 30 seconds. If the vendor cannot demonstrate the control acting, you have an assertion — not evidence.

How Do You Verify Evidence Instead of Accepting Assertions?

A technical due diligence framework for AI agent security grades vendor answers on an evidence ladder with four rungs — assertion, document, configuration, and demonstration — and requires a live demonstration for anything tied to a high-impact action. Certifications such as SOC 2 or ISO 27001 attest to organizational controls, so ask what is in scope, because they do not by themselves show how an agent behaves at runtime.

The eight categories move from knowing the agent to proving what it did:

  1. Scope and inventory — live discovery across all runtimes
  2. Threat assessment — mapped to OWASP Top 10 for Agentic Applications and MITRE ATLAS
  3. Identity and delegation — scoped non-human identities, sub-agent handling
  4. Data privacy and residency — region rules, zero data retention, provider allow-lists
  5. Runtime controls — tool-call inspection, parameter validation, prompt injection defense
  6. Containment and blast radius — kill switch, revocation, microVM isolation
  7. Evidence and audit — tamper-resistant logs, action-level trails separating agent from user activity
  8. Agentic AI governance — policy as code, evaluation gates, supply-chain verification

For each question, grade the answer by the highest rung reached. Require a demonstration (rung 4) for anything tied to a high-impact action: financial transactions, data exfiltration paths, code deployment, infrastructure changes. A document (rung 2) is acceptable for static configurations like region rules. An assertion (rung 1) is never sufficient.

What Does Pricing Look Like When You Factor in Discovery?

Per-agent pricing is marketed as transparent and scalable, yet the Cloud Security Alliance found 82% of enterprises have unknown AI agents. A low per-agent price looks cheap until unknown agents are counted, and those undiscovered agents push the bill past the estimate — making discovery a prerequisite to any meaningful pricing comparison. AgentShield publishes explicit per-agent and per-call tiers as an alternative to hidden enterprise quotes, but the per-agent model is structurally blind to the largest risk: undiscovered agents.

Standard Compute takes a different approach: governance features are priced individually ($199–$399/month each) or bundled in Enterprise at $1,499/month, with no per-agent meter. The Teams platform is free. This aligns cost with control scope rather than agent count — but you still need to know how many keys you’re governing. Microsoft Agent 365 at $15/user/month annual commitment ties cost to human users, not agents, which works for Microsoft-centered estates but leaves multi-vendor agent fleets uncovered.

The hidden implementation expenses matter too: integration overhead, deployment scale, maintenance, patching, and labor costs. Enterprise vs. SMB plan differences often hide compliance-driven pricing premiums. Custom enterprise quotes usually include a floor plus volume with on-prem and SLA terms — five figures a year and up. Model the total cost against your real traffic, not a demo. A metered per-call model can be the cheapest or the most expensive option depending entirely on how chatty your agents are.

Decision Framework: Grading Vendor Answers Before You Sign

Use the evidence ladder as your scoring rubric. For each of the 30 questions across the eight categories, record: the question, the vendor’s answer, the highest rung reached (assertion/document/configuration/demonstration), and whether a demonstration is required for your use case. Weight each answer by the highest rung reached. Require a demonstration for anything tied to a high-impact action.

Create a simple matrix:

CategoryQuestionsDemos RequiredVendor A RungVendor B RungVendor C Rung
Scope & Inventory42
Threat Assessment31
Identity & Delegation42
Data Privacy31
Runtime Controls54
Containment43
Evidence & Audit42
Agentic Governance32

If a vendor cannot reach demonstration rung on your required items, they’re not a fit — regardless of price. The cheapest per-agent quote becomes the most expensive once discovery reveals the shadow fleet. Governance purchases should be ranked by discovery coverage and runtime tool-call authority, not by identity features or per-agent list price, because 82% of enterprises already run agents they do not know about and no identity control can govern an agent missing from inventory.

What’s your threshold for accepting a configuration export instead of a live demo — and which high-impact actions in your environment would you never ship without seeing the control refuse an attack in real time?