Anshu Gupta - Agentic AI Security - What you need to know?
Speaker - Anshu Gupta
Defining the "Agentic" Attack Surface
Concept: moving from "Output" (GenAI) to "Action" (Agents).
Architecture breakdown: Understanding the "Cognitive Loop" (Perception $\rightarrow$ Planning $\rightarrow$ Action $\rightarrow$ Memory) and where vulnerabilities hide in each stage.
Threat Landscape: When Prompts Become Payloads
Deep dive: Indirect Prompt Injection (how reading a malicious website can hijack an agent).
Real-world scenarios: Data exfiltration via API calls, unauthorized purchases, and "confused deputy" attacks where agents are tricked into misusing their permissions.
The OWASP Top 10 for LLMs (Agent Edition)
Focusing on the risks specific to autonomy:
LLM06: Excessive Agency: Granting agents too much power or vague permissions.
LLM07: Insecure Plugin Design: Vulnerabilities in the tools the agent connects to (e.g., SQL databases, email clients).
Zero Trust for Agents: Identity & Access Management
Treating agents as non-human identities.
Strategies for Just-in-Time (JIT) access and scoping OAuth tokens strictly to the task at hand (e.g., "Read Email" vs. "Send Email").
"Hooking Before Hacking": Monitoring & Observability
Why traditional WAFs (Web Application Firewalls) fail against semantic attacks.
Implementing "Thought Monitoring"—analyzing the agent's Chain of Thought (CoT) for malicious intent before it executes a tool call.
Defensive Architectures & Guardrails
Human-in-the-loop (HITL): requiring approval for high-stakes actions (e.g., transferring funds, deleting files).
Sandboxing: executing agent-generated code in ephemeral, isolated environments (e.g., Firecracker microVMs or Docker containers).
Governance & The Future of Multi-Agent Systems
Managing "Swarm" risks: what happens when multiple agents with different instructions conflict?
Essential policy checklist for deploying autonomous agents in the enterprise.