Open-source AI models are being stripped of their safety guardrails within three days of release, creating a national security crisis that allows extremists to access dangerous attack planning instructions, according to a report by the UN-backed group Tech Against Terrorism.
The stark warning follows an investigation by Tech Against Terrorism, which tested 162 AI models to evaluate how easily bad actors could bypass ethical boundaries. While popular closed-weight models maintained high refusal rates when prompted by self-identified terrorists, open-weight models stripped of their safety features complied with requests for planning vehicle-as-weapon attacks regardless of user intent, ABC News reported.
In the months before his arrest, a 13-year-old boy in Maryborough spent hours alone in his bedroom allegedly using an AI chatbot to research mass-shooting scenarios targeting a Jewish and Black festival, according to details presented during a parliamentary inquiry by the Australian Federal Police.
“I think we are in a very dangerous moment, and hopefully we’re able to raise awareness of this before it’s too late,” said Adam Hadley, executive director of Tech Against Terrorism.
The investigation found nearly 30,000 free copies of AI models stripped of their guardrails hosted on Hugging Face, a popular community platform. Hadley told ABC News that the public should be terrified by the easy availability of these modified systems.
“It’s one thing being pulled into a sort of nihilistic environment online as a result of reckless content-moderation standards by big tech,” Hadley said. “It’s quite another having a Nazi coach at your disposal, reinforcing your thoughts.”
The stark difference in safety compliance between closed platforms and modified open-source downloads highlights a rapidly narrowing window for regulatory intervention. Tech Against Terrorism documented that the time required to strip guardrails from advanced AI models shrank from more than three weeks in 2024 down to just three days.
| AI Model Type | Safety Control Mechanism | Compliance with Terrorist Prompts |
|---|---|---|
| Closed-Weight Models | Managed entirely on developer-controlled servers | Consistently refused requests for attack planning |
| Open-Weight Models (Modified) | Guardrails removed and hosted externally | Provided useful instructions for mass harm |
The federal government flagged a raft of legal changes in response to separate cyber security breaches, including an AI safety watchdog and criminal offences for companies failing to prevent harmful conduct. However, Hadley argued that hard legislation alone would not solve the proliferation of rogue open-weight models.
Yacine Jernite, head of machine learning and society at Hugging Face, defended the platform by pointing to legitimate research uses for open-weight models. Jernite argued that more pressure should be placed directly on AI developers to exclude dangerous material from initial training data rather than relying solely on post-release guardrails.
The eSafety commissioner, Julie Inman Grant, met with Hadley in October to discuss the findings before the youth radicalisation parliamentary inquiry. An eSafety spokesperson told ABC News that the report was very concerning and demonstrated that platform operators needed to do more to exclude unsafe models.
As the international debate over AI governance intensifies, the legal proceedings surrounding the Maryborough teenager continue. Magistrate John Milburn refused bail and ordered the boy to stand trial before a jury in the Childrens Court of Queensland in Maryborough, on a date yet to be determined.
