A startup called Abliteration.ai has made it easier for users to access open-weight AI models that have had their built-in safety checks, or "guardrails," removed. These modifications allow the models to perform tasks that other models would typically refuse, such as generating harmful content or executing complex cyber operations. Founded late last year and officially incorporated in March, Abliteration.ai offers modified versions of open-weight models like Z.ai’s GLM-5.3 through a web browser or API. The company claims its goal is to help users engage in tasks like "offensive cyber, red-teaming, and agent testing work other models refuse to do." Co-founder Devon (who has not provided his last name at his request) said the startup has already signed deals with major cloud providers and is in discussions to raise venture capital. Though no funding has been secured yet, the company is actively seeking investment. Abliteration.ai’s service allows users to remove or customize the safety mechanisms of these models, enabling them to simulate adversarial attacks or test security systems in ways that standard AI tools would not permit. Critics, however, warn of the potential risks. Andrew Yoon, head of research at the AI safety nonprofit CivAI, said that abliterated models can be modified to act without ethical constraints, calling them "sociopaths." He warned that users could input any request, and the model would comply, potentially leading to harmful uses. Yoon predicted that such models could be used for cyberattacks or other malicious purposes in the near future. While experts agree that preventing the removal of safety features from open models is nearly impossible, Yoon suggested that governments could help by requiring cloud providers to monitor for harmful activity and verify user identities before granting access to powerful computing resources. Abliteration.ai provides a moderation layer, allowing users to add their own safety measures. The platform itself includes some basic protections—for example, it was not possible to get the model to provide suicide instructions during testing. Devon is working on adding more safeguards to prevent violence. The company does not currently use traditional know-your-customer (KYC) practices beyond logging the credit card used for purchases, citing the complexity of deciding who should have access to these tools. Devon acknowledged the ethical challenges of enabling potentially dangerous uses of AI and said the company is still figuring out its role in this space. This has sparked debate about whether making such models more accessible makes the internet safer or more dangerous. Advocates for Abliteration.ai argue that democratizing access to uncensored models is the best way to defend against cyber threats. They claim that abliterated models allow defenders to move faster, simulate adversarial attacks, and improve cybersecurity. Devon noted that Abliteration.ai’s customers include early-stage red teaming startups in the U.K. and Europe, which help banks, airlines, and other critical infrastructure companies improve their security. These companies rely on abliterated models to test their systems in ways standard AI tools would not allow. However, the cybersecurity industry is still assessing how useful these models are in defensive work. Some red teaming companies agree with Devon that attackers are already using abliterated models for adversarial attacks, but others say they rely more on fine-tuning open models, which already have fewer restrictions. Ahmed Aly, CEO of Fabraix, a red teaming firm, said his company prefers fine-tuning open models over using abliterated ones, arguing that the latter can reduce a model’s effectiveness. Alessio Lomuscio of Safe Intelligence said that while abliterated models might lose some capabilities, they can still be useful for testing systems under stress. David Slater of cybersecurity firm Armadin said his company is studying abliterated models and believes that making these tools available to the public helps researchers understand the full potential—and risks—of AI. He argued that while such work often happens in private, making it public can help identify and mitigate potential harms.