Show Notes
Are we underestimating how the agentic world is impacting cybersecurity? We spoke to Mohan Kumar, who did production security at Box for a deep dive into the threats of true autonomous AI agents. The conversation moves beyond simple LLM applications (like chatbots) to the new world of dynamic, goal-driven agents that can take autonomous actions. Mohan took us through why this shift introduces a new class of threats we aren't prepared for, such as agents developing new, unmonitorable communication methods ("Jibber-link" mode). Mohan shared his top three security threats for AI agents in production: Memory Poisoning: How an agent's trusted memory (long-term, short-term, or entity memory) can be corrupted via indirect prompt injection, altering its core decisions. Tool Misuse: The risk of agents connecting to rogue tools or MCP servers, or having their legitimate tools (like a calendar) exploited for data exfiltration. Privilege Compromise: The critical need to enforce least-privilege on agents that can shift roles and identities, often through misconfiguration.
Guest Socials - Mohan's Linkedin Podcast Twitter - @CloudSecPod If you want to watch videos of this LIVE STREAMED episode and past episodes - Check out our other Cloud Security Social Channels: -Cloud Security Podcast- Youtube - Cloud Security Newsletter If you are interested in AI Cybersecurity, you can check out our sister podcast - AI Security Podcast
Questions asked: (00:00) Introduction(01:30) Who is Mohan Kumar? (Production Security at Box)(03:30) LLM Application vs. AI Agent: What's the Difference?(06:50) "We are totally underestimating" AI agent threats(07:45) Software 3.0: When Prompts Become the New Software(08:20) The "Jibber-link" Threat: Agents Ditching Human Language(10:45) The Top 3 AI Agent Security Threats(11:10) Threat 1: Memory Poisoning & Context Manipulation(14:00) Threat 2: Tool Misuse (e.g., exploiting a calendar tool)(16:50) Threat 3: Privilege Compromise (Least Privilege for Agents)(18:20) How Do You Monitor & Audit Autonomous Agents?(20:30) The Need for "Observer" Agents(24:45) The 6 Components of an AI Agent Architecture(27:00) Threat Modeling: Using CSA's MAESTRO Framework(31:20) Are Leaks Only from Open Source Models or Closed (OpenAI, Claude) Too?(34:10) The "Grandma Trick": Any Model is Susceptible(38:15) Where is AI Agent Security Evolving? (Orchestration, Data, Interface)(42:00) Fun Questions: Hacking MCPs, Skydiving & Risk, Biryani
Resources mentioned during the episode: Mohan’s Udemy Course -AI Security Bootcamp: LLM Hacking Basics Andre Karpathy's "Software 3.0" Concept "Jibber-link Mode" Video CrewAI Framework OWASP Top 10 for LLM Applications Cloud Security Alliance (CSA) MAESTRO Framework
More from Cloud Security Podcast
Is the fear surrounding Anthropic’s Mythos model justified? In this episode, Ashish sits down with Johnny Hand, Global CISO at Trend AI and host of The AI Security Brief. Johnny breaks down why turning on a model like Mythos without a multi-stage orchestration layer is like handing someone the keys to a Ferrari when they don't know how to drive. He talks about the 66,000+ CVE "vulnpocalypse" panic and that only 1% of vulnerabilities are ever actively exploited in the wild. We also examine real-world autonomous threats like "Jade Puffer" (the first documented agentic ransomware path) and walk through 3-step maturity framework for enterprise AI adoption: Visibility, Observability, and Actionability via Agentic Governance Gateways
When companies rapidly migrate to the cloud without strict registrar policies, users often spin up unsanctioned subdomains pointed at third-party services like Vercel or Hugging Face. When those services are spun down but the DNS records remain, it creates "dangling" domains ripe for takeover, a vulnerability that can compromise entire hosted zones, auth tokens, and session cookies. In this episode, Ashish sits down with Ramesh Ramani, Staff Security Engineer at Block, to discuss how his team reduced their domain takeover bug bounty reports from over 80 in a single submission down to near zero. Ramesh breaks down the complex challenge of DNS attribution, explaining why automating DNS cleanup is impossible without first identifying resource owners and consolidating approved registrars. We explore the technical architecture behind Block's stateless, AWS-native automation system, which continuously enumerates their external attack surface and routes vulnerable CNAMEs, MX records, and Elastic IPs directly to the responsible application teams. Ramesh also shares his thoughts on the limitations of AI when evaluating least privilege and parsing logs, emphasizing why security engineers still need deterministic signals.
When an AI coding agent hallucinates a software package name, attackers can simply create malicious packages with that exact name, tricking the AI into downloading malware directly into your environment. In this episode, Ashish sits down with David Gibson, SVP of Strategic Programs at Varonis, to explain why the rapid adoption of AI copilots and autonomous agents is fundamentally a data security crisis. We spoke about how connecting AI to your corporate environment before locking down your data creates massive, unintentional insider threats, allowing employees to instantly discover sensitive information they shouldn't see. We explore the Varonis Atlas framework, which provides a unified control plane to inventory AI systems, implement runtime guardrails, and execute automated red teaming. David also breaks down emerging threats like "hallucination squatting," untrusted agent inputs, and why managing identity groups without understanding data access is like holding a keyring without knowing what doors the keys unlock.
Empowering non-technical users to build production-grade applications requires a proactive, secure-by-design architecture. In this episode, Ashish sits down with Marcus Hallberg and Samuel Kelemen, security engineers at the AI software creation platform Lovable, to discuss how they actively secure AI-generated code and protect creators from supply chain risks. Focusing on solutions rather than alarmism, Marcus and Samuel detail their response to an incident where malicious contractors mimicked AI agent commits. They outline their multi-layered defense strategy: securing a predefined tech stack, utilizing integrated security scanners, and tuning coding agents with strict guardrails to ensure secure output by default. The conversation also covers the evolution of the shared responsibility model in the AI era, the critical importance of foundational hygiene like Git commit signing, and the necessary transition from local "YOLO mode" to secure, sandboxed agent environments. Discover how platforms can leverage AI not just to write code, but to actively assist users in finding and resolving security issues through concepts like "CISO agents."
Hash values, IP addresses, and domains are virtually useless as primary detection mechanisms in the era of AI-driven polymorphic malware and short-lived phishing kits. If your detection strategy is still stuck at the bottom of the Pyramid of Pain, your incident response team is doused in noise while true threats slip through undetected. In this episode, Ashish sits down with Nicole Beckwith, Senior Director of Security Engineering & Operations at Cribl (former Secret Service investigator and Kroger security ops lead), to break down Cribl's open APEX framework. Nicole explains how APEX focuses on behavioral chaining, time-boxing, and clustering over raw telemetry to build high-fidelity detections without needing a thousand individual rules for every MITRE ATT&CK box. We also explore the shift from a "single pane of glass" to a deterministic "single lens" over federated data, why AI agents must be provisioned as true identities rather than unmonitored service accounts, and how to analyze hyper-fast threat archetypes like Anthropic's GTG 1002.