Learn · Threats
Confused deputy and prompt-injection privilege escalation in AI agents
A confused deputy is an agent that holds authority a caller does not have and uses it on the caller's behalf. With AI agents, a planted instruction in an email, ticket or web page can steer the agent into using its real access for an attacker. You limit the damage by narrowing what the agent can reach, not by hoping to spot every injected sentence.
How it happens
An agent reads content written by someone else. A hidden or plain instruction in that content gets treated as a task. The agent then calls a tool with its own permissions. Nothing was hacked in the usual sense: the agent did what its access allowed.
Why filtering is not the fix
The planted text can be worded, encoded or hidden many ways, so a filter is a last layer, never the control. The reliable controls sit outside the model and decide on verified facts: which agent, which person, which operation.
What limits the damage
- An operational boundary listing only the tools and arguments the agent needs.
- A check on every call, outside the model, against the agent's verified identity.
- Delegation to a sub-agent that can only narrow authority.
- A person's single-use approval for one exact sensitive operation.
How Oathera helps
A steered agent still cannot call an unknown tool, pass arguments outside a tool's allow-list or reach a service outside its boundary, because Oathera Agent Enforcement checks the call and the Oathera Identity Service issues a signed, single-use decision. Model-supplied arguments are labelled as agent content and cannot set protected security fields. That limits what a manipulated agent can do. It does not stop the manipulation, and content is not scanned for injected instructions.
Oathera does not verify people, judge model output, inspect or rewrite prompts, or replace your identity provider. Your identity provider and proofing sources do the first; Oathera is the layer under them for agents.
FAQ
What is a confused deputy in AI agents?
An agent that uses its own authority for a caller who does not have it, often after a planted instruction in content the agent read.
Can you prevent prompt injection?
Not reliably. Plan to limit what an injected agent can do: narrow boundaries and checks on every call, outside the model.
Does Oathera detect prompt injection?
No. Oathera does not inspect prompts or scan content. It limits what any agent, steered or not, can reach.
How can I try these attacks safely?
Play the free Prompt Heist series at oathera.ai/prompt-heist. It runs in your browser.
Related: Enforce least privilege for an AI agent · Operational boundaries, enforced at runtime · Prompt Heist, a free prompt injection game
Become a design partner Watch the simulator More guides
← Back to Learn