Risk · R-170 · Agentic & Autonomous Systems
Privilege escalation through agent delegation chains
An agent passes work to other agents with authority that is wider, longer-lasting or deeper than its own, so a delegated task can do what no agent in the chain was granted.
- Severity
- High
- Audience
- Heads-of · Practitioner
- Frameworks
- AML.T0053 · ASI03 · ASI07
- Responsible AI domain
- Lifecycle stage
- Reviewed
- 2026-11-17
What it is
Agents increasingly hand work to other agents: an orchestrator splits a task among workers, which may call further agents and tools. Each hand-off is a delegation of authority. Safe delegation only ever narrows: the receiver gets no more scope than the sender holds, for no longer, and with a limit on how many further hand-offs it may make. The risk is that authority widens or persists along the chain, because credentials are passed whole, tokens are issued without reference to the caller's grant, or nothing records where authority came from. How an agent's own identity is scoped is covered in agent identity and privilege abuse, and the security of messages between agents in insecure inter-agent communication.
How it happens
- A sub-agent is started with the orchestrator's full credentials.
- A worker requests a new token from a service that does not check what the caller holds.
- Delegated permissions have no expiry, or outlive the task that needed them.
- Chains have no depth limit, so an agent can recruit others without end.
- A worker that ran for one user is reused for another without its authority being reset.
Why it matters
Each step can look correct on its own while the chain ends in an action nobody authorised. Because the final action is taken under a legitimate identity, access controls and logs show nothing amiss, and an investigation has to reconstruct the whole chain to find where authority widened.
Warning signs
- Sub-agents share their caller's credentials.
- Delegated tokens carry no record of the grant they were derived from.
- No limit exists on delegation depth or lifetime.
- Nobody can list which agents are able to start which others.
Who owns it
The owner of the orchestrating agent is accountable for what the chain may do. The identity and access management function sets the rules for delegated credentials. Engineering enforces them in the platform that runs the agents.
In practice
A finance assistant delegates the preparation of a supplier report to a worker agent, passing its own token so the worker can read invoices. The token also permits payments. The worker misreads a note in an invoice as an instruction and schedules one. The same pattern generalises to any chain in which authority is handed on whole.
Mitigating controls
Primary controls
- Least-privilege agent identity and permissions Preventive
- Requires each delegation to narrow scope, lifetime and depth. Each agent has its own identity, with permissions scoped to its task and to the user it is acting for, short-lived and reviewed like any other entitlement.
Supporting controls
- Agent action tracing Detective
- Records the authority each sub-agent was given, so the chain can be reconstructed. Every agent action is recorded with the instruction, plan, tool call, result and identity that produced it, in a form that allows the sequence to be reconstructed.
- Agent sandboxing and isolation Preventive
- Contains what a worker agent can reach, whatever authority it is handed. Code execution and tool use by agents run in isolated environments with restricted network, file and system access, so that a failure is contained where it occurs.
- Authority check at the point of execution Preventive
- Confirms when the action takes effect that the delegated authority is still valid. Immediately before an agent's action takes effect, its authority is confirmed again: the delegation, approval, credential and business conditions must all still hold, or the action does not proceed.
Framework and clause references
| Framework | Clause | Title |
|---|---|---|
| MITRE ATLAS | AML.T0053 | AI Agent Tool Invocation Delegated tool invocation is the means by which widened authority is used. |
| OWASP Top 10 for Agentic Applications | ASI03 | Identity and Privilege Abuse Identity and privilege abuse includes authority passed between agents. |
| OWASP Top 10 for Agentic Applications | ASI07 | Insecure Inter-Agent Communication Delegation travels over the channels between agents. |
Sources
- MITRE ATLAS MITRE · framework
Related terms
How to cite
Axiom Verity (2026). Privilege escalation through agent delegation chains. Reference Library, R-170. https://axiom-verity.com/risks/privilege-escalation-through-agent-delegation. Reviewed 2026-11-17.
Found an error or a better source? Report a correction.