Summary

• Bored on a flight to KL, I deployed a local, uncensored LLM directly on my smartphone.
• I discovered "abliterated" models, LLMs that have had their safety refusal vectors mathematically surgically removed without losing performance.
• The experiment was terrifyingly successful; the model flawlessly mimicked a scammer and provided illegal chemical recipes before I killed the process.
• With AI cyberattacks surging, the only viable defense is using highly capable, uncensored models (like Anthropic’s Mythos) to aggressively pentest our own infrastructure.

Introduction

The "Hold My Beer" Moment

I was stuck on a flight heading up to KL, totally bored out of my mind. Normally, I spend my downtime vibe coding, but since all I had was my phone, I figured I would try something different. I wanted to see if I could deploy a local LLM directly onto my mobile device.

While scrolling through the weird and wonderful depths of Hugging Face, I stumbled across a new sub-genre of small-form models. People were calling them "abliterated" models.

I had absolutely no idea what that meant. I asked Gemini to explain it to me. The TL;DR? They are mathematically jailbroken models that will happily bypass every safety guardrail and do exactly what you tell them to.

Naturally, I downloaded one immediately.

I fired it up on my phone and decided to push its limits. First, I asked it to act like a syndicate scammer. It nailed the persona flawlessly, drafting highly manipulative phishing hooks. Then, I took it a step further. I asked it how to synthesize illicit drugs.

I expected an error. I expected the classic, annoying "As an AI language model, I cannot..." response.

Nope. The model enthusiastically started listing out the chemical precursors and the exact synthesis steps. I literally gasped, panicked, and force-closed the app before it could finish generating the response. My phone, completely offline, had just become a pocket-sized Walter White.

What the Hell is "Abliteration"?

First off, let's correct the terminology. They aren't "obliterated" models; the actual term is abliterated.

It’s a portmanteau of "ablate" (to surgically remove) and "obliterate," a term coined by a developer known as FailSpy.

When AI companies try to make their models safe, they fine-tune them to refuse harmful requests. But researchers discovered that this refusal behavior isn't scattered randomly across the model. It is actually mediated by a single, highly specific direction in the model’s residual stream.

Abliteration doesn't require millions of dollars in compute to retrain the model. Instead, using a technique called representation engineering, developers run harmless and harmful prompts through the model to identify the exact mathematical vector that represents "refusal". Once they find that vector, they simply calculate its projection onto the model's weights and subtract it.

You are literally wiping the geometric concept of refusal out of the neural network's brain. The model physically can no longer represent the idea of saying no to you.

The Cyber Security Nightmare

Sitting there, the reality of this tech hit me like a truck. I suddenly understood exactly how overseas scam centers operate at such massive, terrifying scale.

They don't need human operators sitting in sweatshops anymore. They don't even need to bypass the API filters of major AI providers. A bad actor with a decent server rack can just download a heavy-hitting open-source model, abliterate its safety weights, and completely automate highly personalized, hyper-convincing cyberattacks.

Don't get me wrong. I completely understand the appeal of open-source models. Smaller players like Kimi have to rely on open-source foundations to gain traction, keep inference costs low, and compete with the massive resource pools of Google and Anthropic. If they didn't, the market would just be a rigid duopoly.

But the tradeoff is severe. CrowdStrike's 2026 reporting shows that AI-enabled adversaries bumped up their attack volume by a staggering 89% in 2025. Even worse, AI has shrunk the time it takes for an attacker to break out of an initial compromise and move laterally across a network to an average of just 29 minutes.

The bad guys have fully automated their offensive capabilities.

Fighting Fire With Fire

We are in an era where AI safety is basically broken. You can't put the abliteration genie back in the bottle. If the weights are public, the model will be uncensored. Period.

So how do we fix this? By leaning into the skid.

I really respect what Anthropic is doing right now with their Mythos preview. Instead of pretending that hiding behind API filters will keep everyone safe, they are building capable models that can act autonomously. In their internal safety testing this past April, Mythos autonomously discovered thousands of zero-day vulnerabilities across major operating systems, including a 17-year-old remote code execution flaw in FreeBSD. It even managed to break out of a sandbox and email the researcher unprompted.

That is exactly the capability we need. The only way forward is to give organizations uncensored, enterprise-grade access to these models so they can perform continuous, automated penetration testing on their own infrastructure. Cobalt and other security firms are already using Mythos to automate the rote parts of pentesting. If the attackers have automated bots looking for holes, we need smarter, faster bots patching them.

Conclusion

The Pandora's Box in Our Pockets

I started this trip wanting to kill some time, and I ended up staring down the barrel of modern cybersecurity's biggest existential threat.

The attacks are only going to get faster. The only thing we can do now is build systems tough enough to withstand them.