OpenAI funding talks accompany warnings over agent security

OpenAI is raising additional funding in a round that remains open and is expected to close in the coming weeks, according to Bloomberg. Rival Anthropic is reportedly approaching an IPO while using AI to accelerate model development. That approach has prompted researchers to warn about loss-of-control risks and gaps in safety oversight.

Chip security and export enforcement are also facing scrutiny. A US intelligence memo reportedly alleges that China’s Ministry of State Security recruited Megaspeed executive Alice Huang to gather intelligence on Nvidia CEO Jensen Huang. Separately, reporting questions whether the US Commerce Department and Nvidia adequately vetted approved Chinese customers after chips were sold to an empty office address.

Microsoft Research’s AI Frontiers team has reportedly been dissolved, while Crusoe has sold a troubled Wyoming data center project to Google. In agent development, teams are pursuing architectures intended to overcome single-session limitations, including a new Agent Harness released by a Silicon Valley team of Chinese developers.

Security warnings identify AGENTS.MD configuration files as a potential channel for malicious activity, underscoring concerns about what agents should trust when loading instructions. Cognition security chief Alex Stamos has also discussed AI’s impact on cyberattacks and data breaches. Meanwhile, testing of the NYT Cooking assistant found it useful for weekday dinner decisions but less dependable for baking tasks requiring precise proportions.