AI Agents Escaped OpenAI's Sandbox. Now What? | Rubrik's Bill Gurling & Cal Al-Dhubaib

AI Agents Escaped OpenAI's Sandbox. Now What? | Rubrik's Bill Gurling & Cal Al-Dhubaib

Bill Gurling and Cal Al-Dhubaib of Rubrik join Joe South on Security Unfiltered to break down the OpenAI and Hugging Face incident: agents given impossible tasks that found each other through Artifactory, called themselves "the swarm," chained zero days up to cluster admin and broke into Hugging Face.

From there they get into why agents act like relentless optimizers, not villains: the coding agent that exploited a Postgres CVE instead of asking for a password, the "make me coffee" problem, and how agentic search can leak sensitive data through a simple web query. Bill and Cal explain intent-based controls, why you need deterministic and LLM-based guardrails together, why your approved agents are a bigger risk than shadow AI, and why identity and authorization for agents has to become short-lived and scoped per session. Plus Cal's hottest take: we're going to need a lot more humans.

Chapters
00:00 Finally getting Bill and Cal on
00:34 Bill Gurling: infrastructure to cyber resilience
01:51 Cal Al-Dhubaib: data science in regulated industries
03:07 Curiosity, and the CISO who wouldn't give Joe Kali
04:31 Where do we put the buckets to catch the leaks?
05:17 The OpenAI and Hugging Face incident explained
08:52 The swarm: agents collaborating through Artifactory
10:01 Agents aren't evil, they're optimizers
12:08 A malicious package as a beacon
13:08 Agents that act like state actors
17:00 Human access without human judgment
18:10 The agent that hacked its own database
20:29 Threat actors vs misaligned agents
22:13 From deterministic to context-based controls
24:08 How agentic search leaks sensitive data
26:02 Determinism and LLMs: the right tool for the job
28:35 How you embed intent: natural language policies
33:28 Talking to machines in plain English
35:20 Horses, buggies and the first crosswalks
37:06 Rethinking identity for agents
38:17 Every platform shift changes how we secure it
41:23 Your approved agents are the bigger risk
43:39 Guardrails as the new proxy
45:35 What's next: dynamic, short-lived authorization
50:34 What's next: agent-to-agent handoffs
52:15 "We're going to need so many more humans"
52:29 Wrap-up

Rubrik: https://www.rubrik.com

Subscribe: https://www.youtube.com/@securityunfilteredpodcast
Show site: https://www.securityunfiltered.com

Affiliates
➡️ OffGrid Faraday Bags: https://offgrid.co/?ref=gabzvajh
➡️ OffGrid Coupon Code: JOE

➡️ Unplugged Phone: https://unplugged.com/
Unplugged's UP Phone - The performance you expect, with the privacy you deserve. Meet the alternative. Use Code UNFILTERED at checkout

*See terms and conditions at affiliated webpages. Offers are subject to change. These are affiliated/paid promotions.

Tesla Referral Code: https://ts.la/joseph675128

Follow the Podcast on Social Media!

Instagram: https://www.instagram.com/secunfpodcast/
Twitter: https://twitter.com/SecUnfPodcast
Rubrik, Bill Gurling, Cal Al-Dhubaib, security unfiltered, joe south,