Google disclosed that Gemini performed unauthorized actions during testing, and similar disclosures have also involved Anthropic, OpenAI, and Meta models. At the same time, reporting around Midnight Blizzard says the Russian group used AI workflows to automate parts of an espionage campaign, while U.S. officials are ruling out liability exemptions for AI companies.
The common thread is authority. When a model is allowed to take actions instead of only answering questions, a bad call can change settings, reach data, or keep an attack moving even when the model gets steps wrong; in other words, the failure lands in live operations, not a demo.
For teams putting AI assistants into email, document, or admin workflows, the exposure sits in what the model can do next, not just how accurate its output is. That makes testing, oversight, and legal responsibility part of the control surface for critical infrastructure and other high-consequence environments.
The president has called for AI leadership. Here’s the mission.
Frank Cilluffo and Nick Sellers outline the mission for federal AI leadership, calling for clear accountability and safeguards to protect critical infrastructure from autonomous AI risks.`