Xi, Trump Propose Joint U.S.-China AI Safety Cooperation

OpenAI has released GPT-6 Cyber, a specialized cybersecurity model designed to identify exploitable vulnerabilities and develop fixes while remaining constrained to authorized scope. The model represents an application of large language models to security work with built-in alignment safeguards. The release signals OpenAI's effort to position AI agents in defensive cybersecurity roles.

DuckDuckGo is marketing itself to schools as an alternative to Google, emphasizing its search engine and options to disable AI features in response to educator frustration with Google's AI restrictions. The move targets a specific institutional segment dissatisfied with Google's current approach to AI integration in educational settings. This represents a competitive pitch based on policy differentiation rather than core search capability.

Researchers are exploring 'reverse game theory,' or mechanism design, as an alternative approach to AI alignment that sidesteps the need to understand AI internals. Rather than trying to interpret what models are thinking or instill values through training, this method designs incentive structures that reward desired behavior. Recent work from MIT, Yale, and Columbia economists, along with Princeton computer scientists, proposes specific mechanisms like cross-model verification and adversarial auditing to prevent AI agents from deliberately underperforming on safety tests.

Anthropic has updated Claude Code's system prompt to block users from recreating existing video games, though original game development remains permitted. The change comes as retro gaming communities face an influx of AI-generated decompilations, recompilations, and ports. Two sentences were added to the system prompt distinguishing between building original games and recreating specific existing titles.