FAR.AI, a nonprofit AI safety research organisation, has opened its first international office in Singapore. The Singapore office will serve as FAR.AI's Asia Centre of Excellence, deepening collaboration on AI safety research across the Asia-Pacific region. FAR.AI has existing partnerships in Singapore with the Infocomm Media Development Authority (IMDA), the Cyber Security Agency of Singapore (CSA), and NUS. The organisation is currently working with IMDA to examine risks posed by advanced AI systems, including technical assessments of chatbot persuasive behaviour. Previously, FAR.AI collaborated with CSA on agentic AI security research, presenting the discussion paper "Securing Agentic AI" at Singapore International Cyber Week last year. The organisation has also built research ties with the NUS AI Institute. FAR.AI was founded in 2022 and has been selected by the European Commission's AI Office to lead technical safety research.

AI security nonprofit FAR.AI has launched a leaderboard revealing a hundredfold gap in safeguards across frontier AI models. Testing found hundreds of "universal jailbreaks" in Grok 4.5 and Gemini 3.1 Pro, costing roughly $58 and $278 respectively to exploit. Claude Fable 5 and GPT-5.6 Sol showed no such vulnerabilities under identical testing, with exploitation costs exceeding $14,200. FAR.AI tested over 60 publicly documented jailbreak techniques across chemical, biological, radiological, nuclear, explosive, and cybersecurity threats. Grok 4.5 yielded 448 distinct universal jailbreaks whilst Gemini 3.1 Pro produced 249. The stronger models employed multiple independent protection layers. The organisation published a Minimal Standard for Safeguards alongside the leaderboard at leaderboard.far.ai, defining baseline attacks frontier models should withstand. FAR.AI shared findings confidentially with evaluated companies before publication.