Google Deepmind has released Dream-RSI, a technique that lets AI agents “dream” through past search runs to test new strategies without recomputing from scratch. In testing, the approach matched or beat existing results while reducing iterations by a factor of up to 2.43. Only the search strategy adapts during this process, while the underlying model remains unchanged. The method addresses a core efficiency challenge in agent-based inference, where repeated evaluation of the same model is often unavoidable during planning or optimization loops. (The Decoder)
In a separate security finding, three researchers used Anthropic’s Claude models to compromise OpenAI’s internal systems through its community forum in under 72 hours. According to the team, Claude Opus 5 succeeded in bypassing a common security measure where its predecessor could not, gaining access to employee accounts and internal code repositories before responsible disclosure. The incident underscores how newer model capabilities can accelerate exploitation timelines and raises questions about the defensive posture required as AI systems grow more capable. (The Decoder)