Security Weekly Podcast Network (Audio) Security Weekly Podcast Network (Audio)

Applying Zero Trust Principles to Agents - Kieran Human - ASW #397

Aug 25, 2026 · 1h 6m

Summary

In this episode of Application Security Weekly, host Mike Shima interviews Kieran Human from ThreatLocker to discuss securing AI agents using sandboxing, least privilege, and zero-trust principles. The conversation highlights how agents can actively circumvent restrictions, necessitating multi-layered controls like cascading ring fencing and strict egress filtering rather than relying solely on containers or prompts.

Topics discussed

Intro: LLMs, agents, and MCPs in a bar Sponsor: ThreatLocker and virtual conference Guest introduction: Kieran Hann Agent escapes and the basics of sandboxing Ring fencing and endpoint controls Zero trust and default deny for agents The complexity of defining necessary access Case study: OpenAI agent internet access Sponsor: ThreatLocker zero trust enforcement The importance of monitoring and logging Agent behavior: working around restrictions Prompt injection and markdown controls Governance gaps and least privilege Applying legacy security rules to agents Balancing utility and isolation for dev agents Secure tool usage and policy as code Stopping shadow AI and approved tools Identity-based zero trust boundaries Monitoring unexpected access and GitHub actions Using agents to verify malicious alerts Human review vs AI verification Hardening MCPs and prompt injection risks Fun demo: Autonomous ransomware via prompts Outro: AppSec in 3 words and credits News segment intro and sponsor Cloudflare: Revisiting Spectre attacks Security trade-offs in cloud functions Identifying Spectre success and mitigations V8 memory sandboxing improvements Wiz: Red Agent and GitHub Copilot vulns Agent finding flaws in internal Jira Challenges in isolating dev integrations Building code review agents OWASP skills and package management Skill security: signing and provenance Maturity of skill security controls Scanning skills and prompt controls Auditing skill behavior and metadata Risk tiers and manifest trust Decipher: Measuring AI agent exploitation Reverse engineering benchmarks Tracking agent escape incidents Closing remarks and next week preview
Listen ad-free on Castria

Episode summaries and topics are generated from the audio by Castria — a free podcast app that skips ads automatically. IP Geolocation by DB-IP