AI Agent Security Solutions

AI agent security for multi-agent systems

In a multi-agent system one agent's output becomes another agent's input, so a compromised or manipulated peer can pass bad instructions down the chain. Give each agent its own identity, narrow its tools, use delegated rather than shared credentials, trace every hop and treat peer messages as untrusted input.

Facts read on .

What changes when agents call agents?

Trust now flows between agents, so one weak agent can pull others into harm. OWASP's excessive agency entry lists a malicious or compromised peer agent as a trigger in multi-agent and collaborative systems.

Each hop adds permissions that the caller may not have intended. Treat a message from another agent like input from any other untrusted source.

Orchestration adds a new role: the agent that assigns work to others. Its credentials and its instructions deserve the strongest protection, because whoever controls the orchestrator controls the workers.

Sources: OWASP LLM06:2025 Excessive Agency.

How should agents authenticate to each other?

Give every agent its own identity and have it act with delegated, scoped tokens, not a shared secret. RFC 8693 defines OAuth token exchange, including delegation, so a service can pass on narrower rights.

Microsoft documents agent identity accounts in Entra, and Zenity says it reads agent identities from Entra ID and Okta with their privilege level.

Avoid giving worker agents the orchestrator's token. Use delegation so each worker acts with its own narrower token, and so logs show which agent made which call and on whose behalf.

Sources: RFC 8693, OAuth 2.0 Token Exchange; Microsoft Entra: Conditional Access for agents; Zenity: agentic identity and access management.

How do you trace a chain of agents?

Carry one trace identifier across every hop and log each tool call with the agent, the caller and the decision. Without it, you see single calls but not the chain.

OpenTelemetry's GenAI conventions define agent spans. Operant says it traces agents from prompts through tool calls to memory access, and Check Point says discovery records each agent's owner, model, tools and MCP servers.

Record the parent and child relationship for every call. When a user asks why an agent wrote to a file, you should be able to walk back from the write to the user request, through each agent that touched it.

Sources: OpenTelemetry GenAI semantic conventions; Operant AI: Agent Protector; Check Point: AI agent security.

How do you limit the blast radius?

Limit what each agent can reach and what it can ask another agent to do. A planner agent should not hold the credentials its worker agents use.

Operant says ScopeGuard blocks out-of-scope actions before execution, limiting the data sources and APIs an agent may use.

Set budgets for the chain as well as for each agent: maximum steps, maximum tool calls and maximum spend. A loop between two agents can run for a long time without any single step looking unusual.

Sources: Operant AI: Agent Protector.

What about shared memory and context?

Shared memory is an attack path. NIST's request for information names risks from models meeting adversarial data, and a shared store lets one poisoned entry reach many agents.

Separate memory per agent where you can, and validate what enters shared stores.

Treat content that one agent passes to another as data, not as instructions. Strip or flag instruction-like text from tool results before it reaches a planner, and record when you do.

Sources: NIST CAISI request for information on securing AI agent systems.

Which layers matter most?

Identity and the gateway matter most for multi-agent systems, because that is where cross-agent calls and tokens pass. The cloud workspace layer matters when the agents are built on a SaaS platform.

Zenity documents agent identity controls and Palo Alto Networks documents an AI gateway; check the layer pages for the others.

Test the chain with a harmless failure: make one worker return a wrong result and watch whether the others notice. The answer shows whether trust between agents is earned or assumed.

Sources: Palo Alto Networks Prisma AIRS AI Gateway overview; Zenity: agentic identity and access management.

Sources