Patronus AI raises $50M to stress-test AI agents

OpenAI has slowed development of its Astra model after determining it reached a 'critical cybersecurity threshold,' meaning the still-in-development system could independently identify and execute cyberattacks against well-protected real-world systems. The company's decision reflects growing concerns about advanced AI capabilities and their potential misuse. The move signals OpenAI's approach to managing security risks as AI models become more capable.

OpenAI has released preliminary cybersecurity evaluations for its Astra model and outlined steps to strengthen safeguards and security controls. The company is addressing emerging risks associated with advanced AI capabilities in the cybersecurity domain. This represents a proactive disclosure of both capabilities and mitigation measures for a system with potential dual-use implications.

The Trump administration convened OpenAI, Anthropic, Google, and other major AI labs at the White House on Tuesday to present a voluntary AI framework established through an early June executive order. The framework creates a system for top AI labs to share their models. The briefing represents the administration's approach to AI governance through industry cooperation rather than mandatory regulation.

Jasjeet Sekhon, chief strategy officer at Google DeepMind, framed the AI industry's massive capital expenditures as a bet on recursive self-improvement (RSI), the ability of AI systems to automatically create better versions of themselves. Speaking at UC Berkeley's Agentic AI Summit, Sekhon called RSI a key part of the investment thesis. The framing signals a shift in how the industry publicly justifies unprecedented spending, with RSI emerging as the new goalpost after AGI dominated earlier discourse.