Threat Modeling Frameworks Specific to LLM-Powered Agents
Classical threat modeling frameworks fail to capture how LLM agents operate.
Desmond Okwuosa
Staff Writer, Runtime Security
Desmond Okwuosa
Desmond Okwuosa began his career as a systems engineer building container orchestration platforms and transitioned to security research after discovering a class of privilege-escalation bugs in a widely deployed runtime. He covers the defensive side of agent execution environments with a focus on policy enforcement and sandboxing.
8 stories
Minimal Security Controls That Preserve Agent Effectiveness
AI agents need separate defenses for each of their four attack surfaces.
Desmond Okwuosa
Exfiltration Prevention Controls for Tool-Calling Agents
Four control layers stop data theft, but most teams skip the hardest ones.
Desmond Okwuosa
Guardrail Bypass Techniques Against AI Agent Runtime Controls
Attackers now have documented ways to bypass every layer of AI agent security controls.
Desmond Okwuosa
Argument-Level Enforcement in AI Agent Tool Calls
Text-layer content filters miss five critical attack patterns that live in tool arguments instead.
Desmond Okwuosa
Least-Privilege Tool Access Controls for LLM Agents
Rethinking permission scoping and enforcement for AI agents that plan unpredictably across systems.
Desmond Okwuosa
Runtime Guardrails for Browser-Use Agents in Production
Existing security tools miss real-time threats from browser agents operating without human approval.
Desmond Okwuosa
Prompt Injection via Workspace Content in AI IDE Agents
Attackers hide commands in files agents must read to work.
Desmond Okwuosa