How are AI agents compromised?

In your logs, a compromised agent can look like an authorized tool with a valid credential, that’s doing permitted work.In this video, Josh Copeland walks through how attackers can get malicious instructions into an agent and why defenders may struggle to spot them.You'll learn:

• How direct instructions turn an agent into a confused deputy, so whoever can talk to it gets its access.
• How the content the agent reads, even an email nobody opened, can carry an indirect injection.
• How a poisoned tool description can reach every agent that uses the tool.
• Why stealing an agent's credentials can be worse than compromising the agent itself.
• How attackers break the chain between the instruction and the action.