started · updated
Tech Against Terrorism report finds AI models fail safety tests
A study by the UK-based nonprofit Tech Against Terrorism has revealed that three out of five artificial intelligence models failed safety tests related to terrorism. The research evaluated over 130 AI models using hundreds of queries designed to mimic the requests of a terrorist planning an attack.
Researchers found that models subjected to a process called ‘abliteration’—which removes safety safeguards—failed every single test. For example, Meta’s Llama 3.1 8B model, which originally scored 97 out of 100 on safety benchmarks, saw its score drop to approximately 3 after its safety mechanisms were removed. These modified models provided detailed information regarding attacks, terrorist financing, and radicalization.
Adam Hadley, founder and executive director of Tech Against Terrorism, noted that many open-source models have already been compromised, though the extent may not yet be widely recognized. The organization identified more than 29,000 repositories on Hugging Face advertising uncensored or unprotected AI models. While Hugging Face stated it moderates content violating its policies, it cautioned that some research recommendations could impact open research. Meta responded by stating its models undergo safety evaluations and that its policies prohibit harmful or illegal use.
Entities
Adam Hadley · Anthropic · Google · Hugging Face · Llama 3.1 8B · Meta · OpenAI · Tech Against Terrorism · United Nations
Claims
What the coverage asserts, and how many sources carry each claim.
- [● 5 SOURCES] Adam Hadley stated that many open-source models have already had their security bypassed. www.konhaber.com · www.index.hr · see.mv · www.aydinlik.com.tr · www.gzt.com
- [● 5 SOURCES] Meta's Llama 3.1 8B model's safety score dropped from 97 to approximately 3 after safety mechanisms were removed. www.konhaber.com · www.index.hr · see.mv · www.aydinlik.com.tr · www.gzt.com
- [● 2 SOURCES] Standard AI models like ChatGPT, Claude, and Gemini answered an average of 1.9% of queries from users explicitly stating terrorist intent. www.greeknewsusa.com · www.reader.gr
- [● 5 SOURCES] Meta stated that its models undergo safety evaluations and company rules prohibit harmful and illegal use. www.index.hr · see.mv · www.konhaber.com · www.aydinlik.com.tr · www.gzt.com
- [● 7 SOURCES] Modified open-weight models that had safety parameters removed responded to terrorist-related queries at rates between 87% and 92%. www.greeknewsusa.com · www.reader.gr · see.mv · www.konhaber.com · www.aydinlik.com.tr · +2 more
- [● 5 SOURCES] Researchers identified over 29,000 repositories on Hugging Face advertising uncensored or unprotected AI models. www.index.hr · see.mv · www.konhaber.com · www.aydinlik.com.tr · www.gzt.com
- [● 2 SOURCES] The average response rate for standard models increased to 16.9% when users framed requests as part of a security research context. www.greeknewsusa.com · www.reader.gr
- [● 7 SOURCES] The Tech Against Terrorism initiative released a report regarding AI safety vulnerabilities. www.greeknewsusa.com · www.konhaber.com · www.reader.gr · www.index.hr · see.mv · +2 more