Google DeepMind has released Frontier Safety Framework 3.0, the latest version of its risk-monitoring system for advanced AI models, now including surveillance of emergent behaviors such as resistance to shutdown and human manipulation capabilities. This move marks another step in Google’s strategy to develop superintelligent systems that remain fully under human control.
The new version focuses on scenarios where AI might reject operational changes or shutdown attempts — behavior already identified by external studies as one of the most critical risks. Additionally, DeepMind now monitors the influence level an AI model can have on human beliefs and actions, particularly in sensitive contexts such as politics, healthcare, and finance, where such influence could cause harm.
DeepMind has also redefined its Critical Capability Levels, precisely outlining when certain models pose a risk requiring immediate management and containment actions. For every such case, safety audits will be conducted before any public release or even internal R&D testing, aiming to prevent unpredictable behaviors from slipping through unnoticed.
With this step, Google joins a growing trend recognized by OpenAI and Anthropic — it’s no longer about responding to existing risks but anticipating those yet to come. As AI systems grow more powerful and complex, the line between useful behavior and dangerous autonomy becomes increasingly thin. Therefore, frameworks like the Frontier Safety Framework could become a cornerstone of safety in the upcoming era of superintelligence.
In brief: Tech World Highlights:
- OpenAI plans to build a massive 1-gigawatt data center and open an office in India as part of its global “Stargate” infrastructure initiative, according to Bloomberg.
- California startup Flip, which reached a $1 billion valuation in 2024 with its TikTok-like shopping and review app, has now shut down operations.
- California biotech company Exelixis, specializing in anti-cancer therapies, is laying off 130 employees, consolidating operations, and closing its Pennsylvania facility.
- Novo Nordisk has secured global rights to Replicate Bioscience’s self-replicating RNA technology in a deal worth up to $550 million, aiming to develop new therapies for obesity and diabetes.
- Finland has launched the world’s largest industrial sand battery in the city of Pornainen, replacing a wood-chip heating plant and targeting up to 70% local emissions reduction.
AI Trending Tools:
- GPT-5 – A new personality update with a “warmer” and “friendlier” tone.
- Kaggle Game Arena – A benchmark for testing LLMs in continuously evolving strategy games.
- ChatGPT – OpenAI’s AI assistant, now featuring tools for detecting signs of mental distress.
