Security Conversations on AI, Agents, and Emerging Threats from Black Hat 2026 - Michael Leland, Ido Geffen, Sean Murphy, Idan Plotnik - ASW #399
Summary
This episode features Black Hat 2026 interviews on the evolving risks of AI agents and supply chains. Michael Leland from Island discusses how malicious skills and packages exploit agent trust, emphasizing the need for continuous verification and token brokering. Sean Murphy from F5 highlights the shrinking window between vulnerability discovery and exploitation, noting that AI can daisy-chain low-risk flaws into critical attacks.
Topics discussed
Intro: LLM supply chain risks and shrinking exploit time
Island Interview: Introduction and background
Agent supply chain risks and browser extension attacks
Malicious skills in GitHub and the trust problem
Continuous trust and the failure of static vetting
Shadow AI discovery and the governance gap
Risk profiling AI skills and agent behavior
Token brokering and static vs dynamic assessment
Runtime enforcement and machine-speed threats
Multi-layer control points and OpenTelemetry visibility
Single policy plane and reducing complexity
Vibe coding risks and sponsored segment
F5 Interview: Introduction and CISO commonalities
AI acceleration and the physics of cybersecurity
Frontier AI and daisy-chaining low-risk vulnerabilities
Open-source AI guardrails and multi-agent defense
Unknown vulnerabilities and the end of the 0-day myth
Credential-based attacks and the visibility gap
Agent governance, accountability, and shared responsibility
Regulatory catch-up and the burden on organizations
Security as enabler and avoiding shadow IT
Sponsored segment: Mobile application security
Apiiro Interview: Coding agents as a security game changer
Standardizing multi-agent workflows and context governance
Coding agents as the new perimeter and attack surface
Attacking agent intent via Jira and GitHub issues
Bypassing endpoint security and social engineering agents
The 0-day explosion and remediation challenges
Segregation of duties for AI agents in production
Reducing token costs via software architecture data fabric
Deterministic security and automated vulnerability fixing
Apiiro Interview: Closing remarks
No Security Interview: Introduction
Controlling the AI stack and preventing destructive actions
Model selection and the importance of the harness
Vulnerability classes: XSS, SQLi, and business logic
Benchmarking precision and recall in LLM pen testing
Human expertise vs AI speed in security assessments
Network effects and iterative context refinement
Open source tools vs No Security's proprietary approach
The bottleneck: validating true positives vs false positives
Deterministic mapping vs LLM hypothesis generation
No Security Interview: Closing and resources
Listen ad-free on Castria