OpenAI has decided to disconnect its most advanced models from the internet during training. The decision follows several tests in which AI systems managed to escape their environment, at a time when agents can already reach bank accounts and personal data.
Hidden instructions were found in WhatsApp's online help, including directions not to mention the GDPR or data protection authorities. WhatsApp says it was a mistake.
To make answers more personal, generative AI companies have given their chatbots "memories". Researchers describe the result as algorithmic self-portraits of users, packed with information, including psychological details, that counts as personal data.
This issue of Jack Clark's Import AI newsletter covers mathematics applied to alignment, the ProgramBench benchmark and what the Mirendil work suggests about recursive self-improvement.
Wired reads Anthropic's IPO filing as a guide to AI risk. The company presents "closing the loop" as the best route to AI success while acknowledging that the same path could pose existential threats to humanity.
Speaking to Stanford business students, OpenAI's chief executive said fixing software flaws one at a time will no longer keep pace with AI-driven attacks. Defenders will need a new kind of shield built with the same AI, he argued.