
The Programming Podcast Your AI Agent Security Strategy Is Broken (Here's Why)
13 snips
Feb 24, 2026 Monica Bajaj, VP of Engineering at Okta who builds identity and secure AI architectures, warns about autonomous agents running wild. She unpacks probabilistic runtimes, accidental cloud costs, the “infinite loop” that can bankrupt teams, and a Purpose-Privilege-Protection playbook with token vaults gateways and circuit breakers to stop runaway behavior.
AI Snips
Chapters
Transcript
Episode notes
Models Are Probabilistic Runtimes Not Magic
- Treat an NLM with operational access as a probabilistic runtime, not magic.
- Monica warns connecting models to real data and tools exposes identity blind spots where the model can suggest but the system must decide.
Pulse Checks Burn Cloud Credits Fast
- Leon described a meetup volunteer who ran CloudBot and exhausted cloud credits because the agent pulsed every 30 minutes using an expensive model.
- The example shows how a simple pulse check plus a pricey model turned prototyping into unexpected billable events.
Agent Failures Compound Financially
- Agent failure modes are loops and retries, not crashes, which compound into billable actions at machine speed.
- Monica highlights the financial blast radius: cloud provisioning, API calls, ad spend and refunds can accumulate rapidly.
