RR

Ray Redacted

Things Ray Says on Podcasts

Where to Find Them

Ray Redacted has been a guest on Smashing Security (2 times) .

Recently: “Why your AI model could be a sleeper agent” on Smashing Security (November 2026); “I hacked the government, and your headphones are next” on Smashing Security (January 2026).

What They Said

“What Microsoft found was that it is possible for those weights to be modified. And we're not talking about a lot of modification. We're talking about, out of thousands, seven or eight numbers changing.” — Ray Redacted, Smashing Security

Ray is explaining a Microsoft Research paper on backdooring open-weight AI models. Changing a handful of dormant weights can plant a hidden trigger, which gives the episode its "sleeper agent" title.

Smashing Security · 2026-11-19 Permalink → Listen →
Smashing Security Around 34:45 into the episode
Ray Redacted

But until last Saturday,

Graham Cluley

we knew that people could booby trap or backdoor the

Ray Redacted

code that runs in a model, like the Python libraries or the actual code that runs alongside of it. But no one had ever proven that you could actually backdoor a model by changing the weights. Now, just so you understand, the weights is literally a chart full of numbers. Okay. This is a massive chart that looks like nothing but numbers. So you wouldn't be able to necessarily decode the weights. That's how they're trained. But what Microsoft found was that it is possible for those weights to be modified. And we're not talking about a lot of modification. We're talking about out of thousands, seven or eight numbers changing. And if they target dormant ones in what's called the feed forward layer, that's the layer that basically faces the consumer. That's the last layer in the model, so to speak. If they target a dormant one, they can set up a triggered behavior that was not intended in the model to do something like: here's a perfect example. That's a strike of midnight on New Year's Eve.…

Graham Cluley

I would say it is scary, yes. It's a bit like having a sleeper agent, isn't it? Because you've sort of got this piece of code which you're trusting 99% of the time. And what you don't realize is, because no one really understands these things, you know, that the code you're running on your…

Ray Redacted

it's fascinating. You should say the sleeper thing, because if we think about the old cheesy political spy fiddlers, right? They would have a spy or a person that was trained, a Jason Bourne type, and in the middle of his day-to-day activity, the bad guys might have a trigger word like the…

Graham Cluley

I mean, that is when it gets really scary, though, isn't it? Is when it's Agentic AI, which is going to go wrong rather than the chat bot going wrong and telling you something daft. You don't have to tell me if you're not comfortable doing so, Ray, but do you let an AI actually do things on your…

Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.
“The problem is 99% of the people that are using these agents are not doing that. They're defaulting to give it full access to absolutely everything. And by the way, in the case of Muse, this is Facebook. So you do have to ask yourself: would you give Facebook copies of the keys to your house and your car and your children's social security numbers, right? Because millions of people are doing that.” — Ray Redacted, Smashing Security

Graham asked whether Ray lets an AI agent act on his computer. Ray runs his in a read-only container, then points out that most people hand agents full access, including Meta's Muse, which is in the news this week for privacy problems.

Smashing Security · 2026-11-19 Permalink → Listen →
Smashing Security Around 36:03 into the episode
Ray Redacted

…Ermi's agent or Grokbot or Claude or DOT, you may give it full access to your computer, right? Here's a full access to my entire hard drive, et cetera. And suddenly the stakes change. And the next thing you know, your cryptocurrency wallet is just as empty as mine. But mine was stolen by the face.

Graham Cluley

…that is when it gets really scary, though, isn't it? Is when it's Agentic AI, which is going to go wrong rather than the chat bot going wrong and telling you something daft. You don't have to tell me if you're not comfortable doing so, Ray, but do you let an AI actually do things on your computer?

Ray Redacted

100%. Absolutely. Oh, my goodness, Ray. There are ways you can use Agentic AI and be safe. Like, for example, I run Ermes in a container. I use Docker, right? It doesn't have access to read-write activities. If you want to give it financial accounts, there are ways you can do read-only so that it can't write and it doesn't have any credentials, et cetera. But here's the problem, Graham. The problem is 99% of the people that are using these agents are not doing that. They're defaulting to give it full access to absolutely everything. And by the way, in the case of Muse, this is Facebook. So you do have to ask yourself: would you give Facebook copies of the keys to your house and your car and your children's social security numbers, right? Because millions of people are doing that.

Graham Cluley

They've already got that. So why do they need it again? But so, yes,

Ray Redacted

with Agentic, it changes the stakes, so to speak. But in theory, running local models and giving that Agentic ability, if it's properly segmented, if it's got separate VLANs, et cetera, it can be the safest way to do this because you're not transmitting that to the cloud. Microsoft's put a little…

Graham Cluley

So, Ray, come New Year's Eve, right? You're there with your little sherry. You've got your little party hat on. You're waiting for the fireworks and everything. Do you think we should be unplugging our AI agents before midnight? It could be any time, though, couldn't it? It wouldn't necessarily be…

Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.

Collections They Appear In