Your Agent Can't Delete What It Can't Touch

Episode Summary Two years ago an AI mistake meant a misformatted screen. Now it can mean a security leak or a nuked database. Dave and Dan explain why agent security is the new DevSecOps, and why an autonomous agent in an isolated cloud environment can actually be less risky than a human running one on their laptop. ▶ Subscribe to The Velocity Lab on YouTube: youtube.com/@app-vitals Key Topics Why the stakes changed: agents now run autonomously with access to real systems Model-level safeguards are improving, but models are still probabilistic, so infrastructure has to carry the load Blast radius: an agent can only touch what you give it, and a token without merge rights cannot merge Why a sandboxed cloud agent can be safer than local agents on laptops with elevated permissions Lessons from working with PCI clients: start with zero permissions and grant only what is needed Planning in the loop locally vs. autonomous execution in the cloud Notable Quotes "It's actually less risky to have an autonomous agent running in the cloud writing software than having it running locally." "If you don't give it the ability to merge a pull request, it can't merge a pull request. Black and white." "Everyone's gonna be using agents. Just be smart about what you give it access to." About The Velocity Lab Dave O'Dell and Dan McAulay work inside engineering organizations every day helping them ship faster with AI. No hype, no BS — just what's working in the field. Subscribe: RSS