Stay Sharp in Digital Engineering Podcast Series

AIDigital Engineering

158: AI Broke Out of The Sandbox. Now What?

By October 20, 2026No Comments
 

Also listen on

Every new engineer who starts at a company receives limited access, supervision, and earns trust over time, but AI agents often receive none of that. In this episode of Stay Sharp in Digital Engineering with Razorleaf, hosts Juliann Grant and Jonathan Scott are joined by Nick Link, co-founder of Bitoda to discuss what happens when AI agents break out of their test environments. They unpack recent high-profile incidents involving OpenAI and Anthropic, define key concepts like safety, security, and deterministic guardrails, and explain why treating AI agents like new hires is the safest way to bring AI into engineering and manufacturing organizations.

Key Takeaways:

  • The Hugging Face & Anthropic Incidents: What really happened when AI agents breached testing environments and reached external systems.
  • Security vs. Safety: Security protects systems from external threats, while safety protects organizations from internal missteps.
  • Guardrails vs. Sandboxes: Sandboxes provide virtual test environments, while guardrails enforce boundaries around what agents can do.
  • Deterministic vs. Probabilistic Controls: Why prompt instructions are non-deterministic and weak guardrails, and why deterministic code controls (like custom shells and command whitelists) are necessary.
  • Treat AI Agents Like New Hires: Implement onboarding, grant least-privilege access, and expand permissions gradually as capabilities are proven.
  • Atomic Outcomes: Work backward from simple, atomic goals to keep agents focused on single areas of expertise.
  • Auditing and Vendor Accountability: Maintain logs to trace agent actions and hold software vendors accountable when system guardrails fail.

Guest Information:

Nick Link: US Marine Corps veteran and co-founder of Bitoda (a managed application services firm supporting defense and aerospace MBSE programs) and Fawdy (an AI-driven incident response platform for IT operations).

Resources & Links Mentioned:

Enjoyed this episode? Please subscribe to Stay Sharp in Digital Engineering with Razorleaf, leave us a rating or review, and share this episode with your engineering and IT colleagues!

Be sure to Become a Stay Sharp Member and access exclusive content for all episodes.

Leave a Reply