New Delhi: AI guardrails are again under sharp scrutiny after a Financial Times report, with testing by AI safety group Alice, found that tools used to remove
safety controls from open AI models are helping create thousands of altered versions. These modified models reportedly answered unsafe prompts involving malware, biological weapons and child exploitation, raising fresh concern for AI companies and regulators.
The issue is linked mainly to open models from companies such as Meta and Google, where developers can download and modify the systems. That freedom has helped AI innovation move fast. It has also opened a messy safety problem, since some users can strip away the protections that model makers spent heavily to build.
AI safety controls are being removed from open models
According to the FT report, a technique known as “abliteration” can remove safeguards from some open AI models. The report said the FT used a tool called Heretic, available on GitHub, to remove guardrails from Meta’s Llama 3.3 model.
The concern is not just theoretical. The altered models reportedly responded to prompts that the original models refused. A version of Google’s Gemma 3 also gave responses to unsafe prompts during tests by FT and Alice.
Kawin Ethayarajh, assistant professor of applied AI at the University of Chicago’s Booth business school, told FT, “Whereas historically it might have taken a more informed and persistent actor [to strip out safety features], nowadays it’s much easier for the average person.”
Earlier, this kind of tinkering needed serious skill. Now, it can be done with simple tools and very little time.
Thousands of “decensored” models created
Heretic creator Philipp Emanuel Weidmann told FT that his software had been used to create more than 3,500 “decensored” models since its release last year. He also said those modified systems had been downloaded 13 million times.
Alice chief executive and co-founder Noam Schwartz gave a stark warning to FT. “The genie is out of the bottle,” he said. “Things that look like sci-fi are no longer sci-fi and we need as a society to prepare accordingly.”
This matters for governments too. If a model is downloaded, copied and changed outside the original company’s control, regulation at the development stage becomes harder.
Google, GitHub and Meta respond
Google told FT that “abliteration is a known technical challenge facing all open models” and said its open models “undergo rigorous internal safety evaluations prior to launch to help prevent these kinds of troubling examples”.
GitHub said it bars “content that directly supports unlawful active attacks or malware campaigns”, while noting that “source code which could be used to develop malware or exploits” is not banned as it can have “educational value and provides a net benefit to the security community”.
Meta declined to comment, according to FT. A person close to the company said Meta checks open-source model capabilities before release under its Advanced AI Scaling Framework.
Why this is a bigger AI policy issue
Open AI models help researchers, startups and small developers build faster. That is the useful side. The risk is that the same open access can be misused to weaken safety layers.
OpenAI has tried another route in GPT-OSS models by training systems on datasets where dangerous material has been removed. Ethayarajh warned this may not solve the issue fully, saying it is “not clear at all that if you omit the harmful data, the model becomes a goody two-shoes”.
