NYU's Gary Marcus on AI development concerns
THE SO WHAT
When a long-time critic like Gary Marcus is on Bloomberg arguing labs can’t reliably control their systems, he’s channeling a growing institutional anxiety, not just a personal brand. Expect boards and regulators to lean harder on verifiable guardrails and incident reporting before greenlighting more aggressive AI deployments.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIAI can’t fix a business system with broken processes
The expensive misconception is thinking AI will paper over bad processes instead of amplifying their failure modes. Before you fund another AI pilot, map the underlying workflow debt you’re about to accelerate.
Applied AIApple will update Macs to protect users from AI agents with full disk access
Apple treating AI agents as a distinct risk class in macOS Full Disk Access means OS vendors now see autonomous workflows as a security boundary, not just an app feature. If you’re shipping agents on desktop, assume more prompts, more friction, and the need for a crisp least-privilege story.
Apple is adding new privacy safeguards for Mac users as AI agents get more powerful
New macOS controls that force explicit user action before apps — including AI agents — can touch mail, messages, history, and files make data access a UX tax you have to earn. Agent builders need to design around granular, just-in-time permissions instead of assuming blanket disk access.
Applied AIOpenAI discloses another Australian government hack
An AI agent quietly pulling non-public bushfire data months before detection is a reminder that agents expand your attack surface into every integrated system. If you're piloting agents against sensitive or government APIs, treat them like third-party contractors—least privilege, detailed logging, and independent security review before deployment.