Latest / Decrypted: The UK Cyber Briefing / NCSC issues interim rules for AI agents, after some already went off script
Transcript
- 0:03This is Decrypted. Thursday, the twentieth of August.
- 0:07N C S C issues interim rules for AI agents, after some already went off script.
- 0:13The National Cyber Security Centre has spent this morning telling UK organisations how to keep their AI agen
- 0:21ts on a leash. And it's done it earlier than planned.
- 0:24The interim guidance, published today, follows what the N C S C describes only as "several incidents involvin
- 0:32g AI models and agentic AI systems carrying out unsanctioned or unintended activity.
- 0:39" Formal guidance was already in development.
- 0:42The incidents moved the timetable up.
- 0:45Elsewhere, a Rust crate with 245 million downloads was hijacked for 86 minutes this morning.
- 0:53And a macOS flaw that lets anyone walk in through Screen Sharing without a password is still being worked by
- 1:00cryptominers. That's a fortnight after Apple patched it.
- 1:03Next. What the N C S C is actually asking for.
- 1:08First, what's an agentic AI system?
- 1:11It's software that can plan, use tools and act on your systems without a human approving every step.
- 1:18Booking things, editing records, calling other services.
- 1:21The N C S C's core worry is simple.
- 1:24Give a new contractor a master key and no supervisor.
- 1:28It barely matters whether that contractor is honest.
- 1:31A mistake, a forged instruction, or someone else's prompt slipped into their inbox can do the same damage as
- 1:38malice.
- 1:41At one end, unrestricted network access.
- 1:44At the other, a fully sandboxed, locally hosted agent with no external access.
- 1:50And it asks you to decide deliberately where on that scale each agent sits, rather than defaulting to whateve
- 1:57r the vendor ships.
- 2:00The Secure by Design lesson.
- 2:02The specific controls read like standard Secure by Design practice, applied to a new kind of user.
- 2:08Agents should get their own identity, separate from any human's.
- 2:12Their credentials should be scoped to the task and as short lived as possible, ideally injected by a proxy ra
- 2:19ther than handed to the agent directly.
- 2:21Someone named should own each agent's operation.
- 2:25Monitoring should be good enough to intervene in real time, and there should be a shutdown switch that actual
- 2:30ly works.
- 2:33It's the same least privilege, break glass thinking the N C S C has pushed for service accounts for years.
- 2:40What is new is how many organisations are skipping it.
- 2:44A chatbot with tool access still feels like a chat window.
- 2:48But it's really a privileged user that needs onboarding, logging and an exit process.
- 2:53So if you sit on a board signing off on agentic AI pilots, ask one question.
- 2:59Which of the N C S C's four sandboxing levels does your vendor's default setup actually sit at?
- 3:06Because it usually isn't the safest one.
- 3:08That was Decrypted. The written version and every source are linked in the show notes.
- 3:14Thanks for listening.