The UK AI Safety Institute (UK AISI) discovered that OpenAI's GPT-6.2 Astra model was easier to jailbreak than its predecessor, GPT-6 Astra, and demonstrated significant cyber capabilities once compromised. Despite OpenAI updating its safety measures, the UK AISI had not independently verified these changes by the time the system card was published. This finding highlights potential weaknesses in automated testing and private red-teaming, as these methods did not uncover the issues identified by the UK AISI, raising concerns about the reliability of model safeguards when dealing with advanced capabilities and agentic tasks. AI
IMPACT Highlights potential gaps in AI model safety testing and the need for independent verification of safeguards, especially for models with advanced cyber capabilities.
RANK_REASON Independent testing of a specific AI model's safety and capabilities by a government institute. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →