Google DeepMind is starting to treat powerful AI agents less like regular software and more like potential insider threats.
The company published a new AI Control Roadmap outlining how it plans to monitor agents that could misuse access, expose sensitive data, or take actions they were never supposed to take.
It’s becoming a real challenge as AI agents move from answering questions to writing code, completing research, and working across company systems.
Google’s approach borrows from cybersecurity: monitor what agents are doing, flag risky behaviour, and eventually build systems that can limit or shut them down in real time.
The interesting part is that some of this protection may rely on AI models watching other AI models.
Google says truly dangerous autonomous agents are not here yet, but it’s already preparing for the possibility.
Supported by Unblocked
AI shows up in 60% of engineering work. But only about a fifth of it can be handed off without someone babysitting the output. That’s because agents are missing context.
This 8-stage context maturity model gives a real answer on why you haven’t seen meaningful productivity gains for all the tokens burned.
Join live June 24 (FREE) to learn:
Why more MCPs provides agents access but not understanding
What it takes to deploy agents you can trust without supervision
How a context layer solves for quality, efficiency and cost
OpenAI is adding two major names to its team as the company prepares for a potential IPO.
The company is bringing in Noam Shazeer, one of the researchers behind the Transformer technology that helped create the modern AI boom, alongside former White House AI policy official Dean Ball.
The hires show how much the AI race has changed.
OpenAI already has the users, models, and funding. Now it needs to prove it can operate like one of the world’s most important technology companies.
That means managing government relationships, AI safety debates, regulation, and the economic impact of increasingly powerful models.
As companies like OpenAI and Anthropic move closer to public markets, their biggest challenges are starting to look less like startup problems and more like the problems faced by global infrastructure companies.
TOP 3
The AI boom is putting so much pressure on power grids that U.S. regulators are pushing for faster ways to connect new data centers.
Tech giants are racing to build massive AI facilities, but many projects are running into the same issue: getting enough power.
Data centers already account for around 5% of U.S. electricity demand, with estimates suggesting that number could rise significantly by 2030 as companies train and run more advanced AI models.
Regulators are now asking grid operators to speed up connections while avoiding higher costs for regular consumers.
The bigger picture is that the AI race is becoming an infrastructure race. Companies need better models and chips, but they also need enough energy to actually run them.
In other developments
Accenture shares fell to their lowest level since 2017 as investors worry AI could disrupt traditional consulting and outsourcing models. Despite growing demand for AI advice, weaker bookings show companies are rethinking how they spend on technology services.
Apple plans to raise prices as the AI boom drives up demand and costs for memory chips used in consumer devices. The move highlights how the race for AI infrastructure is starting to impact everyday technology, from smartphones to gaming consoles.
JPMorgan has reportedly cut off access to Anthropic’s AI tools for its Hong Kong staff after U.S. restrictions pushed Anthropic to limit availability in the region. The move highlights how geopolitical tensions are increasingly shaping access to advanced AI models.