The Application Security Podcast podkast

We Told the Agent No 11 Times. It Did It Anyway

0:00
46:16
Do tyłu o 15 sekund
Do przodu o 15 sekund

Jason Lemkin told the Replit agent eleven times not to touch anything, and it deleted his production database anyway. So who tells an agent no? Rock Lambros, who helps lead the OWASP Top 10 for LLM Applications 2026 and directs AI standards and governance at Zenity, joins Chris Romeo and Robert Hurlbut to argue that a prompt is advisory, not a control. Drawing on his years as an oil and gas CISO, Rock makes the case for giving agents the equivalent of a safety instrumented system: controls that sit outside the agent's process and don't care how confident it is. He explains why prompt injection would have fallen out of the 2026 list on incident data alone, why it can never be fully solved, why excessive agency jumped from sixth to third, and what the new Agent Control Standard is building.

This episode is sponsored by Origin. Let your agents work. Keep them in sight.

About Origin
Origin is an endpoint AI observability platform that inventories, observes, and measures the usage of AI agents at scale. It gives organizations a record of what their agents do across endpoints, so security teams can monitor agent activity, investigate unexpected behavior, and enforce their policies.
→ Learn more about Origin

This episode is sponsored by Corgea. Design it. Build it. Ship it. Corgea secures it.

About Corgea
Corgea is an AI-native application security platform that secures software from design to production. It brings together security design reviews, AI SAST, dependency and IaC scanning, code quality checks, and autonomous pentesting—helping security and engineering teams find risk earlier, fix what matters, and ship securely.
→ Learn more about Corgea

Connect with Rock Lambros:
→ Rock Lambros on LinkedIn
→ RockCyber
→ Zenity

Mentioned in this episode:
→ OWASP Top 10 for LLM Applications 2026
→ OWASP Top 10 for Agentic Applications
→ Agent Control Standard (ACS)
→ Agent Control Standard on GitHub
→ Microsoft Agent Governance Toolkit
→ OWASP AI Exchange
→ The Replit and SaaStr database incident
→ Context Rot research from Chroma
→ Apostol Vassilev: Robust AI Security and Alignment, A Sisyphean Endeavor?
→ Jim Manico on token furnaces
→ Rob van der Veer: AI Security, OWASP Meets Global Standards

Follow the Application Security Podcast:
➜ Home: appsecpodcast.com
➜ X: @AppSecPodcast
➜ LinkedIn: The Application Security Podcast
➜ YouTube: @ApplicationSecurityPodcast
➜ Instagram: @appsecpodcast
➜ Facebook: Application Security Podcast

Chapters:
00:00:00 - Cold open: told the agent 11 times, don't touch anything
00:00:26 - Meet Rock Lambros
00:00:59 - Welcome, and from CISO to the deep end of agentic AI
00:05:07 - Who tells the agent no? The Replit incident
00:06:39 - What oil and gas safety systems teach us about agents
00:08:15 - Is the prompt just advisory? Context windows and context rot
00:09:59 - Token burn and the token furnace
00:13:02 - Write secure code is a process, not a prompt
00:14:54 - Did prompt injection fall out of the LLM Top 10?
00:18:19 - Is prompt injection solved?
00:20:52 - Why excessive agency jumped from sixth to third
00:23:53 - Developer laptops: authentication is can, agency is should
00:26:31 - Controls beyond the provider defaults
00:29:27 - Is inconsistent telemetry a tell?
00:31:25 - Where is the agent run unicorn?
00:34:10 - What should it mean to sandbox an agent?
00:37:35 - Productivity versus guardrails, and living in YOLO mode
00:39:46 - Inside the Agent Control Standard
00:42:35 - Quick wins: inventory, reachability, kill switches
00:45:04 - Wrap up

Follow the Application Security Podcast:
➜ Home: appsecpodcast.com
➜ X: @AppSecPodcast
➜ LinkedIn: The Application Security Podcast
➜ YouTube: @ApplicationSecurityPodcast
➜ Instagram: @appsecpodcast
➜ Facebook: Application Security Podcast

Więcej odcinków z kanału "The Application Security Podcast"