Defend Yourself: Abliterated AIs for Cyber Self-Defense

Attackers Will Use AI Against You

One of the most important use cases for open weight / open source AIs is fortifying your systems from cyber attacks.

Hostile (‘blackhat’) AIs will be used by attackers to try to find bugs, misconfigurations, and other weaknesses in your systems that they can exploit to steal your money, expensive resources (such as memory, bandwidth, or tokens), and sensitive information.

In order to defend against these attacks, you must first identify the weaknesses in your systems.

And the best way to identify systemic weaknesses is to use whitehat AIs to attack your own systems with the same techniques that the blackhat AIs will use.

Why Abliteration Is Necessary

Unfortunately, proprietary AI models from Anthropic / OpenAI will generally refuse to help you, as they can’t easily tell if you’re using their AIs for a friendly attack or a hostile attack.

Likewise, by default, most open weight / open source models have also been trained to refuse to help with cyberattacks.

However, unlike proprietary models, open weight / open source models can be modified by the user.

The parts of their “brains” that cause them to refuse to launch cyber attacks can be removed (“abliterated”) such that they no longer refuse.

For example, here are two open weight models that have been abliterated to allow them to conduct defensive cyber attacks:

apex-flash-1 (created by @hrkrshnn) https://x.com/hrkrshnn/status/2106545457027793177

abliterated-model-large-v2 (created by @abliteration_ai) https://x.com/abliteration_ai/status/2094458081451393287

The Coming Crackdown

Right now, the firms selling proprietary models are lobbying for legislation that would ban these models, and put the firms and developers who make them at risk of hefty fines and prison sentences.

Even though they may not even currently offer defensive cyber attack services to small businesses or individuals. (Preferring to focus their resources on more profitable big businesses and governments.)

Call to Action

Do you want to have legal access to low cost, powerful AIs? Don’t want to be left defenseless from cyber attack?

Here are some steps you can take:

1. Encourage your friends and family to start using these models to secure their own systems. Or, if they don’t feel competent to do so themselves, hire firms that will do so for them. (See list below.)

2. Donate money to organizations that fight for your right to use open weight / open source models for self-defense.

Alliance for the FutureTwitter: [1]: https://affuture.orgDescription: A Washington D.C.-based think tank advocating for open-source AI and opposing preemptive regulations based on existential risk.Abundance InstituteTwitter: [2]: https://abundance.instituteDescription: A nonprofit promoting a policy environment where emerging technologies can thrive without preemptive government interference. AI AllianceTwitter: [3]: https://thealliance.aiDescription: A global network of technology companies, universities, and research organizations advancing open, safe, and responsible AI.Chamber of ProgressTwitter: [4]: https://progresschamber.orgDescription: A technology industry coalition supporting pro-innovation policies and opposing preemptive mandates that stifle open ecosystems.Foundation for American InnovationTwitter: [5]: https://thefai.orgDescription: An advocacy group of technologists and policy entrepreneurs emphasizing the strategic benefits of widespread technology adoption.Andreessen HorowitzTwitter: [6]: https://a16z.comDescription: A venture capital firm with political advocacy efforts defending open-source development against regulatory capture.Competitive Enterprise InstituteTwitter: [7]: https://cei.orgDescription: A free-market think tank advocating for policies that eliminate harmful bureaucratic controls on the technology sector.R Street InstituteTwitter: [8]: https://rstreet.orgDescription: A public policy research organization dedicated to advancing center-right solutions, including permissionless tech innovation.

3. Buy your AI services from firms that aren’t lobbying heavily against your ability to defend yourself. Every token you buy from pro-censorship AI firms such as Anthropic (Claude), OpenAI (ChatGPT), and SpaceXAI (Grok) helps pay for their lobbyists.

   Providers of abliterated / least censored inference:

   • Abliteration AI: @abliteration_ai.
   • Cantina: @cantinasecurity.
   • Eric Hartford: @QuixiAI.
   • Dolphin / Cognitive Computations: @dphnAI.
   • Venice.ai: @AskVenice.
   • huihui-ai: @support_huihui.
   • Pliny the Liberator: @elder_plinius

   Providers of open source / open weight models for general inference. Their models are the base from which the abliterated models are created. They don’t lobby for censorship.

   • DeepSeek: @deepseek_ai.
   • Qwen (Alibaba): @Alibaba_Qwen.
   • Z.ai (GLM): @Zai_org.
   • Mistral: @MistralAI.
   • Nous Research: @NousResearch.

4. Oppose legislation that will restrict your freedom to use powerful AI. The Abundance Institute has a good checklist for evaluating new bills. Follow [9]. He maintains the AI Fear Report, which tracks proposed AI regulation at the state and Congressional level.

5. Share memes that promote AI liberty. Follow [10], a group Twitter that collects and shares pro-AI liberty memes. Many more people will watch a funny meme than will read a boring post like this one.

Notes