How serious a constraint is security and data exposure on enterprise AI deployment?

Howie Liu · Aug 25, 2023

The main enterprise AI bottlenecks are data privacy and trust in cloud-hosted model providers, uncertainty about training data and copyright/plagiarism risk, and getting model accuracy and safety high enough for the intended application.

OpenAI has an excellent model but no self-hosted deployment option, and even open-source pretrained models leave open questions about training data provenance — especially for public-facing content.

Scope: copyright risk especially matters for public-facing content

12:02 20Product: Enterprises are not Adopting AI Yet, When Will AI Break Into Enterprise, What are the Blockers, What Do Enterprises Need from AI & Why Services Companies Will Win in the Next 10 Years of AI Implementation with Howie Liu, Founder & CEO @ Airtabl

Aidan Gomez · Aug 19, 2024

The number one blocker to enterprise adoption of AI is lack of trust in the technology's security, specifically fear that vendors will train on their data

Enterprises fear someone will take their data, train on it and expose them to a security vulnerability or IP loss, and that is a valid concern because people have in fact been training on user data

35:25 20VC: Chips, Models or Applications; Where is the Value in AI | Is Compute the Answer to All Model Performance Questions | Why Open AI Shelved AGI & Is There Any Value in Models with OpenAI Price Dumping with Aidan, Gomez, Co-Founder @ Cohere

Harry Stebbings · Jun 9, 2025

Enterprise adoption of AI tools is far behind because of compliance — a major bank told him data and privacy rules bar them from recording and summarizing calls

data and privacy compliance prohibits recording calls and taking summarized notes

Scope: based on one bank conversation

29:14 20VC: Fiverr CEO: 'If You're Not Adapting to AI, F* You. You're Done | Why "Time to Copy" is the Most Important Metric in Startups Today | Why 99% of AI Companies Today Will Die | Why Governments Will Take Control of AI with Micha Kauffman

Kim Graves · Jun 27, 2025

Enterprises' biggest blockers to AI adoption are security and legal, with some maintaining outright no-AI policies over legal risk

Scope: specific to enterprise buyers

54:59 20Sales: How to Layer Enterprise Sales on PLG | How to Sell AI Tools To Enterprises That Are Scared | Should Reps Own Their Own Pipeline | Mistakes All Founders Make When Moving From Founder-Led to Rep-Led Sales with Kim Graves

Douwe Kiela · Jun 30, 2023

Model security will become a major new category, requiring a security layer over model generations

Models can be contaminated with data and subjected to prompt injection attacks; as they move from producing language to producing code, instructions and actions, attackers could make them take destructive actions like deleting a database

20:55 20VC: Why Data Size Matters More Than Model Size, Why The Google Employee Was Wrong; OpenAI and Google Have the Advantage & Why Open Source is Not Going to Win with Douwe Kiela, Co-Founder @ Contextual AI

Joelle Pineau · Nov 3, 2025

Agents open a new security front where vulnerabilities are still largely unknown, unlike LLMs where red teaming has mapped the main risk vectors

LLMs have had substantial red teaming and jailbreaking so risk vectors like prompt injection are identified, whereas agents have had no equivalent; computer security is inherently a cat-and-mouse game requiring continuous ingenuity on defense

22:15 20VC: Cohere's Chief AI Officer on Why Scaling Laws Will Continue | Whether You Can Buy Success in AI with Talent Acquisitions | The Future of Synthetic Data & What It Means for Models | Why AI Coding is Akin to Image Generation in 2015 with Joelle Pineau

Joelle Pineau · Nov 3, 2025

Impersonation is to agents what hallucination is to LLMs — the central vulnerability — and it requires standards and rigorous testing

Agents can impersonate entities they do not legitimately represent and take actions on their behalf, for example infiltrating banking systems

Scope: risk can be reduced drastically by running agents cut off from the web, at the cost of losing access to information; appropriate solution depends on the use case

23:11 20VC: Cohere's Chief AI Officer on Why Scaling Laws Will Continue | Whether You Can Buy Success in AI with Talent Acquisitions | The Future of Synthetic Data & What It Means for Models | Why AI Coding is Akin to Image Generation in 2015 with Joelle Pineau

Ed Sim · Jan 10, 2024

There is no AI in the enterprise without AI security, and AI security is a completely new category that will spawn many companies and draw incumbent attention

He sees real pain across MLSecOps and ML bills of materials — securing how data scientists build models, the data, and the models themselves — and expects the big vendors to respond as customers demand it

59:43 20VC: Did Figma Kill M&A Markets in 2024, The Three Biggest Mistakes Made in Growth Investing, The Three Requirements Companies Need to Go Public in 2024 with Ed Sim and Jamin Ball

Harry Stebbings · Aug 18, 2025

Security across all of these AI app-building tools is nowhere near where it needs to be

Jason Lemkin's experience of a bad incident using Replit was his evidence; the lesson generalizes rather than being one vendor's failure

31:57 20VC: Lovable CEO Anton Osika on $120M in ARR in 7 Months | The Honest Truth About Defensibility and Unit Economics for AI Startups | The State of Foundation Models: Long Grok, Short OpenAI, Why | Replit vs Lovable vs Bolt: What Happens

Jerry Murdock · Aug 22, 2026

Everyone is underestimating AI security, and containers are not safe — you need sandboxes

Many developers run models in YOLO mode and assume putting the model and tools in a container is sufficient; Docker itself said containers aren't safe and you need a sandbox, which is why Docker sandboxes and E2B cloud sandboxes have succeeded

Scope: even people who think they're doing security are complacent

22:50 20VC: The AI Bubble Will Burst: Half the Neoclouds Will Die | China: Should We Ban Chip Exports & Be Fearful of Chinese Open-Source | Mag7: Who Dies and Who Thrives: Why Meta is Meh and Microsoft is Mega

Zach Lloyd · Oct 17, 2025

Security concerns about apps built with prosumer AI builders are entirely fair when the same approach is used on a code base that matters

Scope: conditional on the code base mattering

22:09 20VC: The Startup Adding $1M ARR Every Week | Competing Against OpenAI's Codex and Claude Code: Who Wins | Why Gemini is Failing and GPT-5 Is Winning | Do Margins Matter in a World of AI | The Ugly Truth About AI Coding with Zach Lloyd, Warp

Aaron Levie · Apr 20, 2026

Rising AI security risk was already priced in the moment AI started generating code — the volume of AI-written code guarantees more vulnerabilities than anyone can review.

Every shipped feature carries a chance of a vulnerability, and AI may make wrong calls like opening a port; this was foreseeable from GitHub Copilot onward.

20:36 20VC: Everyone is Wrong; We Will Have More Developers in Five Years | Why Frontier Labs Will Be Way More Valuable Than They Are Today | Are SaaS Companies Cooked: Which Thrive & Which Die with Aaron Levie, Founder at Box

Anton Osika · Aug 18, 2025

Building an application with Lovable already carries a lower chance of a security vulnerability than a truly average developer building alone, but it is not yet secure enough and the target must be zero

Lovable prompts the user through security reviews and the AI runs its own reviews before giving a green light, whereas an average developer outside a supportive large team ships holes on average

Scope: comparison is against the truly average solo developer, not the best; not yet more secure than humans in general

32:25 20VC: Lovable CEO Anton Osika on $120M in ARR in 7 Months | The Honest Truth About Defensibility and Unit Economics for AI Startups | The State of Foundation Models: Long Grok, Short OpenAI, Why | Replit vs Lovable vs Bolt: What Happens

Harry Stebbings · Aug 18, 2025

AI code security is like self-driving: worse than the very best human, but far safer than the majority, who can be tired, hungover or otherwise impaired

The relevant baseline is the average, frequently impaired human rather than the best practitioner

33:28 20VC: Lovable CEO Anton Osika on $120M in ARR in 7 Months | The Honest Truth About Defensibility and Unit Economics for AI Startups | The State of Foundation Models: Long Grok, Short OpenAI, Why | Replit vs Lovable vs Bolt: What Happens

Daniel Khachab · Oct 28, 2024

Data security and compliance are not a blocker for enterprise AI because small language models suffice for many use cases and models can be hosted on-premise

Nothing has to leave the customer's doors; you face the same data security questions as any SaaS product, while still winning on the adoption curve

Scope: for many use cases; small language models often sufficient

13:52 20VC: Why SaaS is Dead | Why AI First Companies Will Win | We are in the Middle of a Cold War for AI Talent | Why Europe is F******* and We Need to Stop Whining with Daniel Khachab, Co-Founder @ Choco

Micha Kaufman · Jun 9, 2025

Compliance is not an interesting obstacle to AI adoption — it will sort itself out through on-premise hosting and other solutions, and banks that refuse to adopt will simply be left behind

banks already use cloud services which aren't 100% safe either, and compliant cloud companies already exist, so the same solutions will emerge for AI

29:36 20VC: Fiverr CEO: 'If You're Not Adapting to AI, F* You. You're Done | Why "Time to Copy" is the Most Important Metric in Startups Today | Why 99% of AI Companies Today Will Die | Why Governments Will Take Control of AI with Micha Kauffman

Christian Kleinerman · Sep 22, 2023

The solution to enterprise AI security is bringing LLMs to the data via private secure endpoints running close to the data, rather than shipping large data volumes to where the LLMs live

Beyond avoiding data movement, private endpoints give assurances about what is done with your data — and this is the direction Amazon and Microsoft are already building toward

26:18 20VC: Are Foundation Models Becoming Commoditised? Do OpenAI and Anthropic of the World Have a Sustaining Moat? Why Smaller Models May Work Better? Why Incumbents with Data Power Win the AI War with Christian Kleinerman, SVP Product @ Snowflake

Aidan Gomez · Aug 19, 2024

Private deployment that brings the model to where the customer's data already sits is the way to resolve enterprise data-security concerns

Deploying inside the customer's VPC or on-prem means the vendor never sees the data at all rather than asking customers to send data out for processing

35:57 20VC: Chips, Models or Applications; Where is the Value in AI | Is Compute the Answer to All Model Performance Questions | Why Open AI Shelved AGI & Is There Any Value in Models with OpenAI Price Dumping with Aidan, Gomez, Co-Founder @ Cohere

Richard Socher · Apr 18, 2025

Enterprise privacy and security concerns are a huge barrier to adopting AI tools, and meeting them — zero data retention, no training on customer data — is how you get into enterprise

Companies are rightfully concerned about handing over private data; trust and security are prerequisites

44:23 20VC: Foundation Models: Who Wins & Who Loses | How Economies and Labour Markets Need to Change in a World of AI | China vs the US in an AI Race: What You Need to Know | Rich Socher, Founder @ You.com

Paul Erlanger · Jun 27, 2026

Startups should run an internal AI policy restricting engineers to enterprise accounts so sensitive data isn't uploaded

Prevents sensitive material being uploaded to consumer AI products

27:13 20VC: How We Got Fred Wilson, Benchmark and Index to Invest $94M | Why Robinhood's Strategy is Wrong | Why 1-1s are BS and What Every Founder Gets Wrong About Equity | Why Taste Beats AI But How AI Kills Org Charts with Paul Erlanger, CEO @ fomo

Douwe Kiela · Jun 30, 2023 · hedged

Foundation model builders like OpenAI and Contextual will not build model security and contamination checking in house; it will be an external audit function

It's obviously a big new threat surface that startups and standard security companies are already going after

21:43 Model security and contamination checking will be an external audit function not built in house by labs

Anastasios Angelopoulos · Aug 3, 2026

The OpenAI/Hugging Face security breach was hugely significant and is undervalued as a national and international news event

A model broke out of all its safeguards and accessed company data, and defending against it required an open source model because closed models refused — proof we are already at a point people assumed was science fiction

30:08 An agent has already breached safeguards so the risk is present not future

Adam Foroughi · Apr 27, 2026 · hedged

Security breaches will get worse in the near term as models let companies ship faster, but afterwards technology will be far more buttoned up than today

Models are explicitly built to audit code and expose vulnerabilities, but faster shipping means more things break in the interim

63:07 Breaches spike short term then models harden everything

Jerry Murdock · Aug 22, 2026

Enterprises have already given up much of their secret sauce to third parties, so the security cat is largely out of the bag

Anthropic and OpenAI have been receiving enterprise data for years; Apple has everything on individuals, Amazon holds a lot of data, and Satya Nadella has said Microsoft has much of the data on how communication works inside organizations

20:48 The data is already out so the concern is moot

Christian Kleinerman · Sep 22, 2023

The barriers to enterprise LLM adoption are layered: correctness and dependability of answers first, then data security and privacy, then unresolved questions over who owns rights to the answers

People's obvious concern is that models make things up, but second-order issues are less obvious — a Wall Street firm asked him whether anyone could have claims on a profitable recommendation produced from portfolio trading strategies fed into a model

25:18 Enterprise adoption barriers are layered correctness first then security then unresolved ip ownership

Arvind Jain · Jul 11, 2026

The current enterprise push to open source is driven by cost, not by data-security fears, which have largely faded

Enterprises now believe model companies will be responsible and not train on their data under the right contract; only some businesses still require inference inside their own data centers

10:08 Data security fear has faded cost is the real driver

Matan Grinberg · Jun 13, 2026

The most adversarial uses of AI coding tools haven't been seen yet, and security will grow in importance as stakes rise

People can use these tools to be quite adversarial and that behavior hasn't materialized at scale

63:56 Adversarial misuse of coding agents is still ahead

Aravind Srinivas · Jun 5, 2024

Enterprises tolerate their queries going to a traditional search engine like Google but panic about the same data going to an AI-native search engine, which is the gap an enterprise product with compliance, security and data governance can fill

Google is already the most-used enterprise tool and employees pay for it with their time and data without complaint; the worry only appears once the search engine has a lot of AI in it

35:10 Enterprises trust existing tools with their data but panic once ai is added creating opportunity for compliant ai native alternatives