Startup Abliteration AI removes security restrictions from open-weight AI models
A former Apple engineer named Thomas founded the startup Abliteration AI, which removes security mechanisms (guardrails) from open-weight AI models and makes them available to anyone. This is a response to legitimate queries being refused by models from major companies such as OpenAI and Anthropic.
According to Wired magazine, a 36-year-old engineer named Thomas, an Apple employee, founded a startup called Abliteration AI, which removes built-in security restrictions (guardrails) from the most powerful open-weight AI models and offers the resulting versions freely. Thomas was motivated by frustration that models from major companies repeatedly refused routine work queries — for example, analyzing cyberattack logs or identifying app users with a higher likelihood of paid conversion. To test open-weight models, he built a $20 000 computing cluster at home.
Thomas discovered the process of stripping restrictions from models, referred to in the community as "abliteration," through Reddit, Discord, and X, where a community of engineers has formed that deals with both jailbreaking and modifying the internal parameters of open-weight models. One earlier example is the Dolphin model by developer Eric Hartford, built on Meta's Llama base model and trained without the usual refusal examples.
The article presents criticism according to which OpenAI and Anthropic charge extra for access to models without standard cybersecurity restrictions while also refusing to generate data that could be used to distill other models — pointing to economically motivated reasons alongside safety ones. Thomas and other proponents of "unlocked" models also acknowledge the risks of misuse (e.g. instructions for making drugs or bombs) and defend their approach by arguing that control should rest with the user, not the model provider.
The rest of the article with further details about the Dolphin model and the community was not available.
Why it matters
The existence of readily available "unlocked" versions of powerful open-weight models widens the circle of people who can bypass the security restrictions put in place by major AI companies — whether for legitimate reasons (prototyping, security log analysis) or for misuse. For companies, this means the risk of so-called shadow AI, where employees may use unauthorized models without standard controls outside the oversight of IT and security teams.
Two audiences, two different impacts
What this means
For individuals
For individuals working with AI models (e.g. during prototyping or data analysis), freely available alternatives without the usual query refusals exist, but their use carries a real risk of misuse and potential violation of company policies.
For a business
Companies face the risk that employees may turn to unauthorized "unlocked" models without standard security mechanisms, which increases security and compliance exposure and bypasses the paid tiers of restrictions offered by major AI companies.
Risks and complianceCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.