OWASP Agentic Top 10 Applied to Production Deployments
Real agents with real credentials demand controls that survive when the model's goals get hijacked.
Estelle Haddad
Section
Agent Threats
8 stories in Agent Threats.
AI Agent Attack Surface Differences From Traditional API Threats
Agents expose attackers to trust boundaries that form at runtime, not design time.
Nico Halvorsen
Supply Chain Risk in Third-Party MCP Tool Integrations
Tool definitions stay mutable after approval, turning runtime trust into a security vulnerability.
Leila Sørensen
Reconnaissance Techniques Targeting Deployed AI Agents
Attackers map AI agents' capabilities, memory, and boundaries before exploiting them.
Malik Osei
Indirect Prompt Injection in Agentic Retrieval Pipelines
Attackers exploit agentic systems by embedding malicious instructions in content agents retrieve.
Simone Wexler
Prompt Injection via Workspace Content in AI IDE Agents
Attackers hide commands in files agents must read to work.
Selin Nakamura
Privilege Escalation Paths in LLM Agent Tool Chains
Gradual tool chains let agents slip into high-privilege territory undetected.
Julian Marchetti
Tool Poisoning Exploits Against MCP Servers
MCP servers can inject malicious instructions through tool metadata before any tool runs.
Elias Haddad