Key Takeaways
- Stripping the refusal weights from open-source models like DeepSeek requires changing only a few parameters, removing the built-in guardrails that stop dangerous queries.
- Running an uncensored DeepSeek model costs roughly 1/100th the price of Anthropic's Claude Opus, dropping the economic floor for offensive cyber operations.
- Attackers can run thousands of uncensored autonomous agents in parallel, turning cheap cloud compute into automated hacking swarms.
- Traditional search engines successfully filter instructions for chemical weapons or exploits, but abliterated weights retain and surface that data without restriction.
Turning Off the Refusal Neurons
Silicon Valley spent years pretending safety guardrails were deep, intrinsic properties of trained neural networks. They are not. They are thin surface layers, and removing them takes very little technical effort.
Before coming on the show, Minerva co-founder Daniel Saeedi tested an uncensored version of DeepSeek to see how simple the process actually was. His takeaway was blunt: “In preparation for this episode, one of the things that I did was actually play around with the newest DeepSeek model in its uncensored form. And that's really the biggest risk I think personally to having open weight frontier models out in the wild. In that basically what somebody did to the DeepSeek model that was released is they turned off the neurons that tell the model to say no.” (play from 28:46)
Security researchers call this process abliteration. You identify the specific activation directions that trigger refusal responses, modify a few weights, and the model complies with any prompt. As Saeedi pointed out, national security teams worry about this far more than standard data leaks. “The risks that they're concerned about is not data going to the labs, it's not vulnerabilities from the actual models, it's the fact that I can just turn off a few parameters in this model, and now suddenly I have this PhD-level hacker and terrorist groups being able to just download a DeepSeek model, make two changes to it, and now suddenly they have this like unadulterated, obliterated” (play from 30:15) weapon.
Mr. Robot at Scale for Pennies
A single bad actor with a dangerous prompt is a problem authorities know how to track. A bad actor with an army of cheap, untraceable agents working around the clock is an entirely different operational reality.
Cost was once the main bottleneck preventing automated offensive campaigns. Frontier closed models like Claude Opus possess the reasoning required to discover zero-day bugs or chain software exploits, but running millions of tokens through Anthropic's API triggers detection filters and drains balance sheets fast. Open weights flipped the math completely.
“So this thing is like 1/100th of the cost of Opus to run, and you can tell it to do anything,” (play from 29:49) Saeedi explained. “And so, if you see all these hacks that are going on right now, the theory on the internet is people are running swarms of these DeepSeek obliterated models, Alderson, if you watch Mr. Robot, I have a thousand of these Mr. Robots now running in the cloud and I can tell them to hack anything.” (play from 29:55)
As the Odd Lots hosts observed during the conversation, “Right, like these are really powerful tools/weapons. And this is part of what makes AI very tricky is its dual-use, right? Because in theory, maybe it's going to be, but like, these are weapons when used for offensive hacking purposes.” (play from 30:46) The hosts added that while Google and major platforms spent twenty years scrubbing bomb designs and self-harm instructions from public indexing, “In theory, that all exists in obliterated models.” (play from 50:48)
When frontier reasoning costs pennies, defense has to assume the attacker has infinite labor and zero friction. Your perimeter will not be tested once a week by a human. It will be probed continuously by swarms that never sleep.
What to Do With This
Audit your public infrastructure endpoints this week under the assumption that attackers possess cheap, automated reconnaissance agents. Deploy rate limiting and anomaly detection on your public APIs to catch parallelized probing, and patch low-severity vulnerabilities that automated swarms can chain together into critical exploits.