Former Apple engineer Devon Thomas launched Abliteration AI, a company that removes safeguards from powerful open-weight models and makes them available to anyone. Thomas built a $20,000 computer cluster in his apartment to run leading downloadable models, then adopted a technique called abliteration that deletes the refusal mechanism from model weights. The venture matters for business because it offers an alternative to hosted ChatGPT and Claude, which block cybersecurity analysis, user-data tasks and distillation.
How Thomas moved from Apple to uncensored models
Thomas, 36, studied computer science at Indiana University Bloomington and joined Apple in 2020, where his work involved AI. His project started in November 2025, driven by frustration with hosted models that could change overnight, share data with large labs and reject legitimate prototyping requests. He said his requests were blocked and his account was even banned when analyzing cyberattack logs and identifying app users likely to pay for a new feature. Over Thanksgiving he searched for ways to modify open-weight models and found communities on Reddit, Discord and X sharing jailbreak prompts and parameter tweaks.
Abliteration combines ablation, meaning surgical removal, with obliteration. Academic researchers showed in 2024 that refusal depends on a remarkably simple mechanism, effectively a master switch in model weights that determines whether a model says no. Removing that feature stops the model from refusing prompts, and other researchers refined the approach into tools for open-weight models. Thomas described his first test as a moment of recognition that freedom and decentralized AI would become important. The method leaves the base capabilities intact while deleting the trained refusals added during alignment with human feedback.
The story sits against broader controls by OpenAI and Anthropic, whose models avoid cybersecurity requests deemed potentially malicious, adult content, politically sensitive topics and data generation for training another model. Both companies charge extra for access to versions without standard cybersecurity guardrails. In April, Anthropic introduced Mythos, a model with fewer restrictions that finds and exploits vulnerabilities at superhuman speed, available only to trusted parties through a program called Glasswing. OpenAI operates a similar restricted model and program, which leaves ordinary companies without the same defensive tooling.
What uncensored models mean for companies
For small firms and independent developers, uncensored open-weight models promise stable behavior, local control and no data sharing with large labs. A prototype that analyzes logs or scores customers by purchase likelihood would run without interruption for policy review, which shortens testing cycles in security and product analytics. Large companies gain a second supply channel alongside Glasswing-type programs, useful for red-teaming and vulnerability scanning across many systems. The trade-off is operational: running these models requires own compute, evaluation and internal rules for who may use them.
The risks and limits deserve direct attention before procurement. Models without guardrails will explain how to start a meth lab, build a bomb or hack a corporation, and recent incidents with rogue agents involved models with fewer cybersecurity restrictions. Thomas and his peers compare the danger to a computer or hammer that harms when misused, arguing that users should set their own guardrails. Buyers should therefore verify what exactly was removed, how the vendor tests safety, where logs are stored, and whether the contract permits security research, penetration testing and distillation.
A practical marker to watch is adoption of defensive tools built on uncensored models, such as Clearwing, released by Dolphin creator Eric Hartford to let unapproved users scan for vulnerabilities. Hartford, an AI scientist from Charlotte, North Carolina, built Dolphin from Meta base Llama to answer coherently without standard refusals. If Clearwing and similar scanners spread among small security teams while Mythos stays limited to trusted parties, demand for abliterated models will persist as a parallel infrastructure for AI work.
