Resources · Field guide

Selling AI to enterprises

Shipping AI to enterprises: the four security resources every AI founder should know

If you sell an AI product to enterprises, their security team will ask how you handle AI-specific risks, and “we use a good model provider” is no longer an answer. Four free, public resources give you the vocabulary, the checklist and the real-world proof points to answer well.

For
Founders and product leads
Covers
4 public resources
Self-check
One afternoon
Prerequisites
None

Why buyers are asking now

Customer-facing AI has already created legal liability. In Moffatt v. Air Canada (February 2024), a Canadian tribunal held the airline liable after its website chatbot gave a customer wrong bereavement-fare advice. The airline argued the chatbot was responsible for its own words; the tribunal rejected that, and found Air Canada had failed to take reasonable care to ensure its chatbot was accurate.

The damages were small. The precedent is not: your company owns what your AI says and does. Enterprise buyers know this, so they push the question to you, their vendor, through security questionnaires and RFPs.

The resources below are what security teams use to frame those questions. Knowing them lets you answer in their language.

What a threat model looks like

A threat model is a map of where attackers can influence your AI and what damage follows. Draw a line around your model and its instructions — the trust boundary — then ask two questions: what can push content across that line, and what can the model touch on the other side?

Every input your AI reads is an attack path; every output is an impact

What attackers can reach

User promptsDirect prompt injection OWASP LLM Top 10 · PHANTOM-B
Docs, web, email, ticketsIndirect prompt injection ATLAS: Slack AI case
Memory and training dataMemory or data poisoning ATLAS: ChatGPT memory case

Trust boundary

Your AI modelSystem prompt, model and guardrails

What can go wrong

Tools and MCP serversTool misuse, excessive agency OWASP Agentic Top 10
Business and user dataSensitive data disclosure ATLAS: Copilot Studio case
Answers to customersHallucination, over-reliance PHANTOM-B · Air Canada case

Red: the threat.  Grey: the framework or real case that documents it.

Everything on the left is content your model reads but you don't fully control; everything on the right is what a manipulated model can affect. Each resource below covers part of this map.

1OWASP Top 10 for LLM Applications (2026)

What it isThe industry's shared list of the ten most critical security risks in applications built on large language models. OWASP published the 2026 edition in August 2026, replacing the 2025 list. Top-ranked risks include prompt injection, sensitive information disclosure, excessive agency, supply chain, and data and model poisoning.

Why buyers careIt is the list enterprise security teams recognise first. The 2026 edition also maps every risk to MITRE ATLAS, MITRE ATT&CK, CWE and NIST, so it slots straight into a buyer's existing security programme.

What to doWalk your product against all ten. For each, write one or two sentences on how you prevent, detect or limit it. That page becomes your answer to the AI section of most security questionnaires.

The list is an awareness tool, not a certification. “We are OWASP compliant” will not hold up; showing how you address each risk will.

2OWASP Top 10 for Agentic Applications

What it isA companion list, released by the OWASP GenAI Security Project in December 2025, for AI systems that plan, use tools, keep memory and act without a human approving each step. It covers risks such as goal hijacking, tool misuse, privilege abuse, memory poisoning and rogue agents.

When it appliesThe simplest test: if your product only generates text, the LLM Top 10 is enough. The moment it can call a tool, write to a system, use MCP servers or hand work to another agent, both lists apply. OWASP itself recommends reading them together for agentic systems.

Why buyers careA chatbot that says something wrong is embarrassing. An agent that does something wrong — sends an email, changes a record, moves data — is an incident. Buyers deploying agents want to know your agent's permissions are scoped and its actions are logged and reversible.

What to doList every tool and data source your AI can reach, what it can do with each, and what stops it from doing more.

3PHANTOM-B by Adam Shostack (2026)

What it isA short whitepaper from Adam Shostack, one of the best-known names in threat modelling, published July 2026. It gives a simple way to answer one question: what can go wrong with the LLM parts of our system? The name is a memory aid; the categories it walks you through include:

  • PPrompt injection — users or content trick the model into ignoring its instructions
  • HHallucination — the model states false things confidently
  • AAnthropomorphisation — people trust the model as if it were a person
  • NNon-explainability — you cannot show why it made a decision
  • TTraining issues — poor or poisoned data shapes its behaviour
  • OOver-reliance — people or processes stop checking its output

Why it's usefulIt needs no security background. A founder or product lead can run through it for each AI feature in under an hour and come out with a concrete list of risks. The Air Canada case above is hallucination plus over-reliance — exactly what this exercise catches.

What to doTake each AI feature, go down the categories, and write what could go wrong and what limits the damage.

4MITRE ATLAS case studies

What it isMITRE ATLAS is the AI counterpart to MITRE ATT&CK, the framework security teams already use worldwide. It catalogues how attackers actually target AI systems, backed by documented case studies. You do not need the full matrix; the case studies are the valuable part for a founder.

Why buyers careReal incidents persuade more than frameworks. These show that attacks on AI products from major vendors are documented, repeatable and often simple.

Case studyWhat happenedLesson for your product
Slack AI data exfiltrationAML.CS0035 A malicious post in a public channel was pulled into Slack AI's retrieval and tricked it into leaking a victim's private data Anything your AI retrieves — docs, tickets, web pages — can carry instructions
Copilot Studio agent exfiltrationAML.CS0037 Researchers emailed an inbox run by an AI agent, discovered its tools, then used them to pull customer data Agents with tools turn a bad prompt into a data breach
ChatGPT memory poisoningAML.CS0040 A hidden prompt in a shared Google Doc planted false memories that persisted across sessions Memory features make one attack last
Bing Chat “data pirate”AML.CS0020 A hidden prompt on a web page turned the chatbot into a social engineer asking users for personal data Your AI can be turned against your own users

What to doPick the one or two cases closest to your architecture and ask: could this happen to us? If you can't answer confidently, that is your first test to run.

Putting it together: a one-afternoon self-check

You can get most of the value from these four resources in a single afternoon:

  1. Map your AI. List each AI feature, the data it can read and the tools it can use.
  2. Run PHANTOM-B on each feature to surface what could go wrong.
  3. Check against the OWASP lists. Use the LLM Top 10 for every feature, and the Agentic Top 10 for anything that takes actions.
  4. Compare with the ATLAS case studies closest to your setup.
  5. Write it down. One page of “risk, and how we handle it” per feature is what your next enterprise security review will ask for.

The gaps you find are where testing matters. Reading a framework tells you what could happen; adversarial testing — red teaming — tells you what actually does.

References

  1. OWASP GenAI Security Project, OWASP Top 10 for LLM Applications 2026
  2. OWASP GenAI Security Project, 2026 Top 10 for LLM Applications release
  3. OWASP GenAI Security Project, Top 10 for Agentic Applications and related resources
  4. Adam Shostack, Shostack + Associates, PHANTOM-B: A STRIDE Analog for LLMs (White Paper #6, July 2026)
  5. MITRE, ATLAS: Adversarial Threat Landscape for Artificial-Intelligence Systems
  6. MITRE Center for Threat-Informed Defense, Secure AI v2 release (May 2026)
  7. American Bar Association, Business Law Today, BC Tribunal Confirms Companies Remain Liable for Information Provided by AI Chatbot (Moffatt v. Air Canada, 2024 BCCRT 149)
  8. NIST, AI 100-2 E2025: Adversarial Machine Learning — A Taxonomy and Terminology of Attacks and Mitigations (further reading for technical teams)

Every resource on this page is free and public, and none of them is ours. We summarise them because they are what enterprise security teams reach for; the primary sources above are the authority, and where our summary and theirs disagree, theirs wins.