Sign up to save your podcasts
Or
Share Vulnerability Management in the Age of AI
Share to email
Share to Facebook
Share to X
Share to email
Share to Facebook
Share to X
Or
Vulnerability Management in the Age of AI
The security community has been sounding alarms about AI infrastructure for two years. BadHost, a critical authentication bypass discovered in May 2026 inside one of the most widely used Python frameworks on the internet, is what that alarm sounds like when it goes off. Dan Fernandez returns to AI Spy to talk about what actually happened, why the official severity rating understates the real risk, and what the broader pattern of 40-plus CVEs against MCP implementations in 2026 alone tells us about how AI infrastructure is being built. This is not a technical deep-dive. It is a practitioner briefing on the vulnerability surface that every organization deploying agentic AI is now sitting on, whether they know it or not.
Show Notes
Dan Fernandez, cybersecurity practitioner and returning AI Spy guest.
Dan has spent his career at the intersection of threat intelligence and enterprise security infrastructure. He returns to AI Spy for his second appearance having previously covered AI as an attack tool. This time the conversation flips: AI infrastructure itself is the attack surface.
In this episode:
- Vulnerability management 101 and what fundamentally changed once AI became load-bearing infrastructure inside organisations, not just a layer on top of them
- Why BadHost (CVE-2026-48710), a flaw in the Starlette Python framework used by Netflix, Uber, Microsoft and hundreds of thousands of other projects, carries a severity rating that most practitioners consider a polite fiction
- How MCP servers aggregate credentials across every service an AI agent is connected to, and why that makes a single compromised server a portfolio-level exposure event
- Claude Mythos found over 10,000 vulnerabilities through Project Glasswing and missed this one. What that tells us about where AI-assisted security tooling is strong and where it has genuine blind spots
- 40-plus CVEs against MCP implementations in 2026 alone. Whether BadHost is an outlier or a pattern in how the AI infrastructure layer is being built
- What a defensible enterprise posture actually looks like for organisations deploying AI agents, AI-generated code, and MCP-connected systems right now
Links and References: BadHost - CVE-2026-48710 Starlette Host-Header Auth BypassDisclosing the BADHOST Vulnerability in Starlette – OSTIF.orgDan Fernández | LinkedIn
View all episodes
By Anurag MohapatraThe security community has been sounding alarms about AI infrastructure for two years. BadHost, a critical authentication bypass discovered in May 2026 inside one of the most widely used Python frameworks on the internet, is what that alarm sounds like when it goes off. Dan Fernandez returns to AI Spy to talk about what actually happened, why the official severity rating understates the real risk, and what the broader pattern of 40-plus CVEs against MCP implementations in 2026 alone tells us about how AI infrastructure is being built. This is not a technical deep-dive. It is a practitioner briefing on the vulnerability surface that every organization deploying agentic AI is now sitting on, whether they know it or not.
Show Notes
Dan Fernandez, cybersecurity practitioner and returning AI Spy guest.
Dan has spent his career at the intersection of threat intelligence and enterprise security infrastructure. He returns to AI Spy for his second appearance having previously covered AI as an attack tool. This time the conversation flips: AI infrastructure itself is the attack surface.
In this episode:
- Vulnerability management 101 and what fundamentally changed once AI became load-bearing infrastructure inside organisations, not just a layer on top of them
- Why BadHost (CVE-2026-48710), a flaw in the Starlette Python framework used by Netflix, Uber, Microsoft and hundreds of thousands of other projects, carries a severity rating that most practitioners consider a polite fiction
- How MCP servers aggregate credentials across every service an AI agent is connected to, and why that makes a single compromised server a portfolio-level exposure event
- Claude Mythos found over 10,000 vulnerabilities through Project Glasswing and missed this one. What that tells us about where AI-assisted security tooling is strong and where it has genuine blind spots
- 40-plus CVEs against MCP implementations in 2026 alone. Whether BadHost is an outlier or a pattern in how the AI infrastructure layer is being built
- What a defensible enterprise posture actually looks like for organisations deploying AI agents, AI-generated code, and MCP-connected systems right now
Links and References: BadHost - CVE-2026-48710 Starlette Host-Header Auth BypassDisclosing the BADHOST Vulnerability in Starlette – OSTIF.orgDan Fernández | LinkedIn