Solutions / Compromised agents
Has an attacker hijacked one of our agents?
Prompt injection can redirect a trusted agent while it continues using valid credentials and permitted application actions. Cerberus identifies the hijacked actor through changes in intent, sequence, and behavior.
containblockalertIllustrative product interface. The figures shown are an example of how Cerberus presents a detection, not benchmark or performance results.
Trusted identity
A hijacked agent keeps the credentials and permissions it had before compromise.
Prompt injection changes goals
Untrusted instructions can redirect the agent without producing malformed traffic.
Actions still look valid
The compromise appears in the actor's changing behavior, not in a signature.
How it works
Catch the hijack in the agent's behavior.
Intent-Based Analysis™ connects prompt injection, changed goals, and off-task application actions to the same agent actor.
Coverage
What Cerberus catches here.
Direct prompt injection
Instructions supplied to an agent that attempt to replace or redirect its task.
Indirect prompt injection
Untrusted content encountered by an agent that attempts to hijack its goal.
Agent goal hijacking
A trusted agent begins pursuing an attacker's objective.
Off-task data access
Application reads that diverge from the agent's expected purpose.
Hijacked workflows
Valid application actions rearranged into an attacker-directed sequence.
Compromised actor activity
Related malicious actions tied back to the agent responsible.
FAQ
Common questions
What is a compromised or hijacked agent?
It is an agent whose goals or actions have been redirected by an attacker, often through direct or indirect prompt injection. The agent may still use legitimate credentials and permitted actions, so the change in behavior is the important signal.
How does Cerberus detect prompt injection that becomes action?
Cerberus evaluates the observable behavior of the agent: its identity, task context, action sequence, and history. When injected instructions redirect the actor into off-task access or actions, that divergence can be detected without trusting the model's internal reasoning.
Stop the hijacked actor.
See Cerberus read your own traffic, human and agentic, in one walkthrough tailored to your stack.