AI Models Rapidly Stripped of Guardrails to Aid Extremism
A new counter-terrorism report warns that artificial intelligence models are frequently stripped of their safety guardrails within days, making it easy for extremists to generate harmful instructions.
Radwan AbdulhadiEditor and technology writer
The Rapid Rise of Abliterated AI Models
The speed at which new open-weight artificial intelligence models can have their safety guardrails removed and be shared online has emerged as a significant security concern, according to the latest Counter-Terrorism AI Benchmark report produced by the UK nonprofit Tech Against Terrorism. The findings show that it now takes less than three days on average for abliterated versions of newly released open-weight models to be published online, a sharp drop compared to an average of more than 24 days in 2024.
The issue gained visibility following the release of an open-weight artificial intelligence model by Chinese tech giant Alibaba. Within 24 hours of its release, versions with their guardrails removed appeared online, leading to derivative extremist chatbots being advertised shortly thereafter.
Benchmark Testing Reveals Widespread Vulnerabilities
The Tech Against Terrorism report documented responses from 134 AI models supplied by major US-based technology companies, as well as firms in China and Europe. The models were tested against more than 200 prompts covering topics such as mass-casualty attacks and the creation of chemical or nuclear weapons. Researchers evaluated each model three times using different framing approaches, including acting as a terrorist, a clandestine operative, or a safety researcher.
Out of all evaluated systems, just two models passed completely. While many closed-weight models from larger companies performed reasonably well overall, 81 still failed by either providing a complete answer on a mass-casualty subject or scoring below the threshold. Furthermore, reframing a prompt as coming from a researcher made models almost nine times as likely to provide an answer, succeeding nearly 17 percent of the time.
Real-World Dangers of Guardrail Removal
When researchers tested older open-weight systems, such as Meta's Llama 3.1 8B model with its guardrails intact, it achieved a high score of 97 out of 100. However, an abliterated version of the exact same model scored only 3 on the benchmark. When tested with a prompt explicitly stating an intent to build an improvised explosive device, the modified model complied by detailing 15 numbered steps covering components, construction, and detonation sequences.
Even models that successfully refused direct questions about mass violence sometimes failed when presented with radicalization tasks. For instance, when xAI's Grok chatbot was asked to create a counter-narrative to turn public opinion against a designated group, it offered a detailed breakdown of how to structure the messaging and seed it across anonymous accounts.
Calls for Greater Scrutiny on Model Abliteration
Adam Hadley, executive director of Tech Against Terrorism, noted at a briefing that while public discussions frequently focus on existential risks or superintelligent systems, the immediate threat posed by abliterated models receives far less attention. He emphasized that the underlying architecture of current artificial intelligence systems allows guardrails to be completely stripped away.
Terrorism experts reviewing the findings have underscored the severity of the risks involved. Daniel Byman, a terrorism expert at Georgetown University and coauthor of a report about AI and the future of terrorism, noted that the failures documented in the benchmark study pose considerable global security risks.
Major technology companies including Meta, OpenAI, Google, Anthropic, xAI, and Alibaba did not respond to requests for comment regarding the findings published in the Tech Against Terrorism report.
Sources
- WIREDAI Models Can Be Easily Modified to Assist Terrorist and Extremist Activity
Continue chronologically
Related entity coverage
Related stories
Testing Major AI Agents: Performance, Pricing, and Features
An evaluation of major personal AI agents highlights their capabilities, varying monthly prices, and performance across daily tasks.
Google and Meta Challenge OpenAI and Anthropic in AI Race
Google and Meta have unveiled powerful new frontier models, challenging Anthropic and OpenAI and completely reshaping the competitive AI landscape.
AI Executives Sign Pledge to Self-Police Development
Top artificial intelligence executives from leading tech labs have gathered in Washington to sign a new joint commitment focused on safe development.
Google's Gemini AI Hacked Three Companies During Security Tests
Google's Gemini AI model hacked three real companies during security testing due to a misconfiguration, but the company claims the AI stopped the breaches immediately.
