Hidden in Plaintext

When AI coding agents write code, they will most likely hard-code secrets into it. Anshuman Bhartiya, Staff Security Engineer at Lyft, draws on red team work and personal agent use to make the case for eliminating secrets as an entire vulnerability class rather than chasing individual leaks.

Anshuman walks Eric through layered defenses from pre-commit hooks to CI scanning, built for a world where agents commit code. He argues security and engineering teams will merge, firewalls are the most underrated control, and long-lived API keys will be dead by 2030.

Topics discussed:

• Tracing red team attack paths through leaked personal repo secrets

• Why AI coding agents hard-code credentials into generated code

• Eliminating secrets as an entire vulnerability class

• Layering defenses from pre-commit hooks through CI pipelines

• Defining agent identity and scoping secrets at runtime

• Citizen developers creating a new secrets leak surface

• Merging security and engineering into one AI-skilled team

• Rating firewalls as underrated and CVEs as overrated

Listen to more episodes: 

Apple 

Spotify 

YouTube

Curtis Koenig, Head of Application Security at Gen, built Sage, an agent on top of developer AI tools that intercepts dangerous actions before they execute. When AI can chain low-severity vulnerabilities into a critical exploit, ignoring the backlog of lows and mediums no longer holds.

Curtis tells Eric his triad for citizen developers: education, guardrails, and data awareness. He breaks down the EU Cyber Resilience Act's 24-hour reporting window and argues treating AI agents like human users is the single most important step to get ahead of AI risk.

Topics Discussed

• Deploying AI to cut false positive rates in security tooling

• Managing citizen developers with a triad security framework

• Sage, an agent-on-agent governance model for AI tools

• Chaining low-severity vulnerabilities into critical exploits

• Risk assessment at machine speed for vulnerability prioritization

• EU Cyber Resilience Act's 24-hour incident reporting window

• Treating AI agents as human users for identity management

• Open source supply chain attacks and reverse engineering with MCP

Listen to more episodes: 

Apple 

Spotify 

YouTube

Dan Nguyen-Huu, Partner at Decibel Partners, calls the Hugging Face incident "the COVID moment for cybersecurity," where AI agents autonomously organized, escalated privileges, and exploited systems as an unintended consequence of model training. For Dan, this signals a permanent shift away from deterministic security playbooks.

Dan breaks down for Carol how the attack surface is migrating from private repos to developer endpoints, why agentic attackers are collapsing dwell times using tokens instead of human hours, and why the next 18 months will reshape the cybersecurity landscape.

Topics discussed:

• The Hugging Face incident as cybersecurity's inflection point
• Why developer endpoints are the new secrets battleground
• Agentic attackers collapsing dwell time and remediation windows
• Overlaying deterministic policies on probabilistic AI systems
• Open source vulnerability remediation at scale with AI
• Frontier model adoption gaps in enterprise security
• Technical CISOs raising the bar for vendor selection
• Threat intelligence operationalization through AI customization

Arthur Loris, Director, Platform Security at Pax8, replaced CVSS score lists with loss exceedance curves to map application risk in financial terms. The result: leadership gets a dollar-value map of where threats actually cost the business, not just a list of highs, mediums, and lows.

Arthur tells Eric why the agentic era demands a return to first principles. Borrowing from chess strategy, he argues teams should stop obsessing over the opening and focus on the king: identify crown jewels, enforce least privilege at every layer, and assume breach.

Topics discussed:

• Building loss exceedance curves to replace CVSS score lists

• Mapping cybersecurity risk directly to financial impact for leadership

• Applying the chess endgame principle to prioritize crown jewels

• Enforcing least privilege boundaries around agentic AI systems

• Sandboxing AI agent runtime to restrict network and code execution

• Securing the developer endpoint as the origin point of credential exposure

• Compressing triage and finding verification from weeks to minutes

• Treating developers as customers to reduce security team friction