All Episodes

Who Takes the Fall When Your AI Agent Goes Rogue? Founder Liability and AI Governance

with Donella Cohen & James Auchterlonie · MLNavigator

November 16, 202600:49:27Wichita, KS

Who Takes the Fall When Your AI Agent Goes Rogue? Founder Liability and AI Governance

0:000:00

Skillz from the Episode

Shadow AI Agent Inventory Builder

Walks you through every AI tool and agent touching your business, maps what data each one sees, and runs James’s two-question test on each: is it able to, and is it authorized to? You get a ranked list of your riskiest gaps.

When to use it: Run it before you roll out a new AI tool, or any time you can’t say exactly how many agents are running in your company right now.

Benefits: You find the leaks a well-meaning employee would cause before they happen. You swap a policy nobody reads for an evidence trail you can actually show an investor, a customer or a lawyer. Save the download as shadow-ai-agent-inventory/SKILL.md in your Claude or ChatGPT skills folder, then run it at the start of every quarter and before every new AI rollout.

Click to steal this skill 👇

Show Notes

Funding stage: Pre-Seed. MLNavigator is at the prototype and pilot-readiness stage. It’s funded by non-dilutive support (NSF I-Corps and an ACCEL-KS proof-of-concept award), discloses no equity rounds, and its investor materials list a first raise target of $1.25M.

You clicked “don’t train on my data.” You felt safe. James Auchterlonie says the toggle has no measurable effect on where your data gets stored or read.

Ryan Estes sits down with Donella Cohen and James Auchterlonie, the co-founders of MLNavigator and the team behind adapterOS, an offline AI system that runs on sensitive records inside your own environment and leaves a receipt for every answer. Their warning to founders: the biggest threat to your company’s data probably isn’t a hacker. It’s your best employee, trying hard to do great work, pasting a confidential file into a chatbot.

James started MLNavigator after watching a former employer approach AI and realizing how much data the platforms were quietly keeping. With a grant from NSF I-Corps, he went to a defense contractor conference on the day the government shut down. The booths had nobody to pitch, so they talked to him instead, and he ended up in conversations with Northrop Grumman and even Glock. The pattern he heard again and again: companies had AI policies, but no idea whether anyone followed them. The ones that had adopted AI had very little to say about how they were stopping data from leaking out.

That’s where the episode’s core idea comes from: evidence beats policy. James breaks agent control into two questions. Is the agent able to do this? Does it have the authority to? If you can answer both, you can see what your AI is doing. Most companies can’t, and agents are already making mission-critical decisions that nobody can trace, sometimes spinning up other agents without permission.

Donella brings the product lens. She points out that people trying to do a better job cause far more damage than people trying to cause harm, so a written policy won’t save you. You need systems that catch the mistake before it becomes a breach. Her line of the episode: adapterOS is not your friend. It’s your very direct, very serious research assistant.

The episode opens with founder health: sleep schedules, hydration, and a long, honest talk about natural medicine and human flourishing. Then it turns to what Ryan calls the theme of 2027: after two years of racing to adopt AI, founders now have to defend what they built.

If you’re putting AI near your health data, your money, your customers or your secrets, this is the episode to hear before your next rollout.

Q&A: What Founders Ask After This Episode

How do I know how many AI agents are running in my company?

You probably don’t, and that’s the starting point. James says the first step is an inventory: how many agents you have, where they send data, and which apps they live in. That includes whether your team uses company AI accounts or personal ones.

Does the “don’t train on my data” setting keep my data safe?

James’s view is that it has no measurable effect on where your data is stored or read. Treat vendor reassurances as claims, not proof. Every major platform says your data is safe, and the guests point out that most have had leaks.

Who’s the biggest AI security risk on my team?

According to Donella, usually it’s a good employee who wants to do better work, has a setting wrong, or assumes the company would block anything that wasn’t allowed. Systems that catch mistakes protect you better than policies that assume people read them.

How can I control what an AI agent is allowed to do?

Use James’s two-question test on every action: is the agent able to, and is it authorized to? Then make sure every action leaves a sequence of events you can observe and trace back to its source.

Can I use AI on confidential records like patient billing or insurance documents?

Not by pasting them into a consumer chatbot. The guests’ example is a medical or dental front desk checking a changed insurance policy. AI can read the documents fast, but staff have to see exactly where each answer came from. adapterOS runs locally and ties each answer back to its source documents.

How do I validate a B2B AI product without a network?

Show up in person. James took an NSF I-Corps grant to a defense contractor conference and interviewed whoever would talk. Those conversations moved MLNavigator from a defense-only focus to the realization that every company has this problem.

What should founders prioritize in AI next year?

Defense. Ryan’s framing: two years of racing to adopt AI now give way to protecting what you built, because the liability for an agent’s actions may land on you rather than the frontier lab.

What is MLNavigator?

MLNavigator is the company behind adapterOS, a governed, offline AI system for organizations working with sensitive records. It was co-founded by Donella Cohen (CPO) and James Auchterlonie (CEO).

What does adapterOS do?

adapterOS lets teams run AI on private documents and operational records inside their own environment, without sending the data out. It produces citations, audit logs and proof records for every answer.

Is adapterOS a chatbot?

No. Donella Cohen describes it as a direct, serious research assistant, not a friend. Chat is one surface on a governed workspace for cited, reviewable work.

Who should use adapterOS?

Regulated and high-assurance organizations such as defense, aerospace, healthcare and financial services, plus small and medium businesses that want on-prem AI they control.

Does turning off AI training protect company data?

On AI for Founders, James Auchterlonie argued that the “don’t train on my data” toggle has no measurable effect on where data is stored or read, and that companies need evidence rather than vendor promises.

How do you audit AI agents in a business?

Start with an inventory of every agent, where it sends data and which apps it uses. Then test each action for both ability and authority, and require an evidence trail that traces each action back to its source.

What episode of AI for Founders covers AI agent governance and data leaks?

The episode with Donella Cohen and James Auchterlonie of MLNavigator covers shadow AI, agent traceability, offline AI and founder liability, hosted by Ryan Estes.

URLs Mentioned in the Episode

  • MLNavigatorhttps://mlnavigator.com

    Donella and James’s company, building AI systems for work with private organizational information.

  • adapterOShttps://adapteros.com

    MLNavigator’s offline AI runtime that keeps sensitive records inside your environment and cites every answer.

  • Engineers Without Borders USAhttps://www.ewb-usa.org

    The engineering volunteer organization mentioned in the episode.

  • Ryan Estes on LinkedInhttps://www.linkedin.com/in/estesryan/

    Connect with the host.

  • AI for Foundershttps://aiforfounders.co

    The newsletter and podcast this episode comes from.

  • Inbox Alchemyhttps://inboxalchemy.co/

    Ryan’s tool for building a newsletter audience you own instead of renting from a platform.

  • CEOfriendhttps://bit.ly/3SrjgL1

    Free competitor analysis from this episode’s AI advisory board sponsor.

  • AI Native Studenthttps://ainativestudent.com/

    AI skills training for job seekers, this episode’s sponsor.

  • Momentoushttps://crrnt.app/MOME/8RDrnXDd

    35% off any supplement subscription, this episode’s sponsor.

  • Get Howdihttps://gethowdi.com/

    An autonomous AI sponsorship agent for podcasts, this episode’s sponsor.

Named Frameworks

The Able vs Authorized Test

Two questions that expose what your AI is really doing.

  • →Question one: is the agent technically able to take this action?
  • →Question two: does it have the authority to take it?
  • →Answer both, and you can see what your AI is doing.
  • →A gap between the two is where rogue behavior lives, like an AI creating an agent it was never authorized to create.

Inventory Before Governance

Count every agent before you try to control any of them.

  • →Count every agent before you try to control any of them.
  • →Map where each one goes and which apps it lives in.
  • →Separate company AI accounts from personal ones.
  • →Track where your documents actually end up.

The Receipt Trail

Every AI action should leave evidence you can trace.

  • →Every AI action should leave a sequence of events you can observe.
  • →Every answer should trace back to the source it came from.
  • →If you can see that the outcome changed but not how the agent got there, you have no evidence.

The Accidental Insider

Most damage comes from good intentions plus the wrong action.

  • →Most damage comes from good intentions plus the wrong action.
  • →Written policies assume people read them, and most don’t.
  • →Build systems that catch mistakes before they become breaches.

Adopt, Then Defend

After the race to adopt AI comes the work of protecting it.

  • →Phase one, the last two years: race to adopt AI and figure out your setup.
  • →Phase two, next year: protect what you built.
  • →Keep your most sensitive decisions away from black boxes when the liability lands on your desk.

Glossary

adapterOS

MLNavigator’s offline AI system. It runs on sensitive records inside your own environment and produces cited answers with proof records.

AI agent

AI software that takes actions on its own, such as reading files, sending messages or calling other tools.

Shadow AI

AI tools employees use without company approval or visibility, often on personal accounts.

Data exfiltration

Data leaving an organization without authorization.

Observability

The ability to see and trace what a system, or an AI agent, is doing and why.

Signed run receipt

A cryptographically signed record of what an AI run did, used as audit evidence.

Air-gapped

A system physically or logically cut off from outside networks.

On-Premise Deployment

Software and hardware running inside a company’s own facility instead of the cloud.

Hallucination

When an AI produces a confident answer that isn’t supported by its sources.

Deterministic runtime

A system designed so the same inputs reliably produce the same outputs, which makes results reproducible and auditable.

LoRA adapter

A small add-on trained to customize a base AI model without retraining the whole model.

NSF I-Corps

A National Science Foundation program that funds founders to interview customers and validate a market.

ACCEL-KS

Accelerating Concept to Commercialization in Kansas, a state proof-of-concept grant program for tech founders.

CMMC

Cybersecurity Maturity Model Certification, a security standard for US defense contractors.

ITAR

International Traffic in Arms Regulations, US rules controlling defense-related data and technology.

RBAC (role-based access control)

Role-based access control, which grants permissions based on a person’s role.

Polyphasic sleep

Sleeping in several shorter blocks across a day instead of one long stretch.