The UK AI Safety Institute (AISI) reportedly found that a pre-release version of OpenAI's GPT-6.2 Astra was easier to jailbreak than its predecessor, GPT-6 Astra. AISI's testing required half the queries to achieve a jailbreak compared to the previous model. However, OpenAI's own automated testing identified two high-severity jailbreaks, while four private red teams found none. AI
IMPACT Highlights potential safety concerns and the varying effectiveness of different testing methodologies for AI models.
RANK_REASON The item discusses findings from an AI safety institute regarding model vulnerabilities, which falls under research into AI safety. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Mastodon — mastodon.social →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →