Share from 0:00
0:00
/

Default Deny for AI Agents

ThreatLocker Chief Product Officer Rob Allen on why the business now wants agents running, why we can't patch our way out, and governing everything an agent calls.

Watch above, or listen on Spotify and Apple Podcasts. Timestamps are at the bottom.

A friend in Ireland asked Rob Allen to help him set up OpenClaw, and Rob told him there was no way. Very quickly after, Rob’s customers were asking the opposite question, and now their businesses are pushing security to let tools like Claude Code and Cowork run.

In this episode, I sit down with Rob Allen, Chief Product Officer at Resilient Cyber partner ThreatLocker, to dig into what default deny looks like when the software you’re governing is an AI agent that calls a dozen other programs on your behalf.

Rob has more than 20 years in IT, starting as a system administrator.

He joined ThreatLocker in 2021 as VP of Operations for EMEA, when it was about 25 people, and became CPO in 2023. The company is now over 750 people.

We start with AI as what Rob calls “the most two sides thing there is right now,” including the wave of AI slop exploits after PHP2Shell. From there we get into vendor questionnaires, locking AI tenancies to a couple of IP addresses, and why we can’t patch our way out.

The middle of the conversation is ring fencing. Rob walks through how he used it against the PHP2Shell remote code execution, how to translate a business use case into restrictions, and why you run policies in monitor mode before you enforce them.

We close on ThreatLocker’s new Zero Trust AI Control, released the week before we recorded. That includes the blind spot Cowork creates when it spins up its own VM, and what it takes to govern every process an agent like Claude launches.

We recorded on Oct 26, the same day Anthropic filed its S-1 and cited Cowork’s adoption. Agent governance now reaches the entire workforce, well beyond the engineers security teams are used to working with.


Thanks for reading the Resilient Cyber Newsletter! Subscribe for FREE and join 22,000+ readers to receive weekly updates with the latest news across AppSec, Leadership, AI, Supply Chain, and more for Cybersecurity.


Prefer to listen?

Apple Podcasts

Spotify

Please be sure to subscribe and leave a rating/review, as it truly helps the show!


We discuss:

  • Rob’s path from about 25 people at ThreatLocker in 2021 to more than 750 today

  • AI as a double-edged tool, PHP2Shell, and the script kiddies feeding vulnerabilities into AI tools for POCs that mostly don’t work

  • Why a vendor questionnaire still has to be verified against what the tool actually does

  • The OpenClaw conversations that went from customers asking whether it was running on their machines to pressure from the business to let agents run safely

  • My security versus utility trade-off, and why enterprise subscriptions alone don’t define what an agent can do

  • Restricting Claude and OpenAI tenancies to one or two IP addresses so a stolen token is useless

  • Deny by default versus whack-a-mole on steroids, and patches that break things so people stop installing them

  • Ring fencing PHP so it runs but can’t call other programs

  • PrintNightmare and the gap between exploitable and weaponizable

  • Running policies in monitor mode before you enforce them

  • An MCP server over ThreatLocker’s logs that found one customer approving almost every request with learning mode plus elevation

  • AI as a force multiplier for the data security products already collect

  • Zero Trust AI Control, locking Claude to the company account on company machines, and the Cowork VM blind spot

  • Why in-tool permission prompts turn into an okay button

  • Controlling the full process tree an agent launches

  • Anthropic’s S-1 and agent governance for the entire workforce

Key Takeaways

Assume it’s full of holes

“Assume all of the software that you use every single day is full of holes. Because it all is.”

“So control what it can do, where it can go, what data it can access, where it can go on the internet, what other applications it can call. And as I said, the majority of these vulnerabilities, while still existing, will cease to be exploitable.”

I’ve been writing all year that we’re not going to patch our way out of this. As I told Rob, there are simply too many things being found, too many disclosures and too many things that need to be patched, and that assumes a patch is even available when you need it.

Rob added the part teams feel every month.

Patches break things, and people who get burned stop installing them, so they fall further behind and inevitably run into an incident. That is a vicious cycle, and constraining what software can do is one of the few ways out of it.

Exploitable is different from weaponizable

“So the difference between a vulnerability that is exploitable and weaponizable and a vulnerability that is exploitable but then can’t do anything is really profound.”

Rob tested PHP2Shell himself. Most of the public AI-generated exploits didn’t work, but his team found that remote code execution worked in one particular setup, and a ring fencing policy that lets PHP run without calling any other program shut it down.

That distinction matters more now than it ever has. Vulnerability discovery and reporting have gone up exponentially this year, and nobody’s security budget, team or headcount has gone up with them.

From panic to pressure

“So whereas the first conversations were people panicking going, I don’t want this thing running on my machines, the conversation has now evolved to the point whereas we’re now being put under pressure by the business to allow these things to run and we want them to run safely.”

I’ve been framing this as a security versus utility trade-off, and Rob’s OpenClaw story fits it exactly. Utility will almost always win over security, because people want the value and want to be more productive, and more utility brings more risk with it.

When customers asked whether OpenClaw was running on their machines, the conversation was short for default deny customers. If they never explicitly allowed it, it isn’t running.

In Rob’s words, “You don’t have to decide something’s bad in order to block it.”

Govern the behavior

“It’s all very well and good controlling one thing you need to control the 10, 15, 20 different things that that thing may also call.”

“It was like Claude calling PowerShell, calling Git, calling Bash, calling PowerShell again, calling Claude again.”

Applying ring fencing to behavior, beyond network connections and traffic, is what I like most about this approach. Agents take actions, call tools and do things on your behalf on the system, so a policy that only covers the agent’s own binary misses most of what it does.

ThreatLocker’s AI Control policies carry the ring fence down that whole chain. Rob also covered the Cowork case, where the agent may spin up its own VM and the host has no visibility into what happens inside it.

The okay, okay, okay button

“If you’ve got a yes or no prompt, your average user is just gonna click yes, yes, yes, yes, yes. I mean it might as well just be an okay, okay, okay, okay button for the amount of times that people will actually click no.”

In-tool permission prompts are the control most organizations are leaning on today. Rob’s point is that a prompt inside the tool leaves the decision with a user who will click yes. That is why ThreatLocker extended file and internet controls down into the agent’s sandbox.

Let AI read the logs

“It’s not like a human being going through, you know, 500 pages of logs trying to figure out what’s relevant, what’s not relevant. You can literally say, go through this, figure out what’s relevant, and tell me about it.”

One MSP asked ThreatLocker’s unreleased MCP server a few plain questions. It surfaced a customer that almost always approved requests by putting the machine into learning mode and pairing it with elevation, which Rob called “completely defeating the purpose.”

This has been one of the more exciting trends this year among vendors and practitioners alike.

Doing this kind of analysis used to be impractical, and getting value from a capable product took a lot of human time and toil. Now you can ask plain language questions and run checks on a schedule.

The whole workforce has agents now

Anthropic filed its S-1 the day we recorded, and it says Cowork’s adoption outpaced Claude Code’s at the same point after launch. In security we’re used to working with engineers and developers, and now we’re talking about the entire workforce with access to systems and data.

Most organizations have little to no visibility into what these tools are doing in the enterprise, and no governance or control over it. Rob’s case for enforcing outside the agent, starting in monitor mode, is a practical place to start closing that gap.


Thanks again to Rob for coming back on the show.

Where is your team on agents today, still blocking them or under pressure to let them run? Let me know in the comments.

Rob Allen

LinkedIn

ThreatLocker | ThreatLocker on LinkedIn

Zero Trust AI Control | Launch announcement

Related on Resilient Cyber

Timestamps

0:00 Cold open

1:20 Welcome and Rob’s six years at ThreatLocker

2:18 AI is the most two-sided thing in security

4:25 Vendor questionnaires versus what a tool actually does

6:01 OpenClaw, from a friend’s request to customer worries

7:31 Security versus utility and enterprise subscriptions

8:03 Locking AI tenancies to one or two IP addresses

9:31 Deny by default and whack-a-mole on steroids

10:34 Ring fencing PHP2Shell

11:12 Assume all software is full of holes

12:55 Turning a business use case into restrictions

14:18 PrintNightmare, exploitable versus weaponizable

15:00 Monitor before you enforce

16:20 Metrics and an MCP server over ThreatLocker’s logs

17:51 Learning mode plus elevation

19:39 AI as a force multiplier

20:29 ThreatLocker AI Control and company-only Claude

21:34 Cowork, sandboxes and the VM blind spot

23:14 The okay, okay, okay button

24:10 Claude’s process tree is a forest

25:57 Governance for the whole workforce

Discussion about this video

User's avatar

Ready for more?