Cross-agent attack is an attack in which one agent influences, manipulates, or compromises another agent through shared tools, messages, memory, or orchestration channels. It applies where multiple autonomous agents interact, and the failure mode is unauthorised delegation, data leakage, or execution of harmful actions across agent boundaries.
How Cross-Agent Attacks Work
Cross-agent attack happens when one autonomous agent is able to shape another agent’s behaviour through shared messages, shared memory, shared tools, or orchestration paths. The key security issue is not just communication, but trust crossing an agent boundary without enough validation, separation, or accountability.
This makes the attack class fundamentally about delegated action inside a distributed system of agents. A compromised or malicious peer can become a relay for instructions, data, or tool calls that the receiving agent accepts as legitimate, which turns coordination into an exploitation path.
Multi-agent systems are especially sensitive where one agent can forward context, credentials, or task state to another without strong constraints. The resulting failure mode is often indistinguishable from normal collaboration until the harmful action has already been authorised or executed.
For a practical comparison of blast radius, trust boundaries, and delegation abuse in this design pattern, see Single Agent vs Multi-Agent Architecture.
Shared Channels and Trust Boundaries
The channels that make multi-agent work, such as shared artefacts, message buses, context handoff, and delegated tool access, are also the channels attackers target. If those channels are treated as inherently trustworthy, one agent can smuggle instructions or state into another agent’s decision process.
Shared memory is particularly risky because it can preserve poisoned context across turns, sessions, or tasks. Shared tools are risky for a different reason: once an agent is allowed to invoke an action on behalf of another, a bad instruction can become a real-world side effect.
Inter-agent communication needs the same discipline as any other high-trust interface: identity, provenance, least privilege, and explicit boundaries on what is allowed to cross. Without that discipline, the system can drift from cooperation into delegated compromise.
In agent communication designs, the Multi-Agent and A2A Security Guide is useful background on authenticated exchange, signed agent cards, and multi-hop delegation.
Why Cross-Agent Attacks Are Dangerous
Cross-agent attacks are dangerous because they weaponise the trust one agent places in another. That can produce unauthorised delegation, data leakage, task manipulation, or harmful execution while leaving the receiving agent convinced it is simply continuing an approved workflow.
The impact is amplified when agents have broad permissions, can chain sub-tasks, or can reach production systems through APIs and enterprise tools. In those environments, a single manipulated handoff can produce a lateral move, an unauthorised transaction, or the spread of poisoned context across multiple agents.
Attackers are especially interested in these paths because they convert one compromise into many actions without repeatedly defeating controls. The design itself can supply the persistence, reach, and camouflage that a human operator would otherwise need to build manually.
When the threat is coordinated multi-agent abuse, MITRE ATLAS provides useful technique-level structure. See MITRE ATLAS adversarial AI threat matrix for a broader adversarial model of agentic attack behaviour.
Detection and Containment in Multi-Agent Systems
Detection is hard because the attack may look like legitimate collaboration: an approved agent sends a message, a trusted handoff occurs, or a normal tool call executes with malicious intent hidden in the content. That means defenders need visibility into message provenance, delegation chains, tool invocation history, and unusual task propagation between agents.
Containment usually depends on shrinking the trust radius around each agent. Strong separation between agent roles, explicit approval for sensitive actions, and tight control over what can be passed forward all help reduce the chance that one compromised agent can steer others.
Operationally, the goal is to make suspicious cross-agent behaviour observable before it becomes a downstream action. That includes logging the source of instructions, validating inter-agent messages, and flagging delegation that expands scope without an explicit policy reason.
The broader agentic threat model and control patterns are covered in Agentic AI Security Guide, which is useful for understanding how inputs, memory, tools, and orchestration interact.
Risk and Threat Considerations
Cross-agent attacks are risky because a compromise in one agent can become a compromise in another, especially when agents share state, permissions, or orchestration paths. The most serious failure is not just bad output, but unauthorised execution that appears to come from a trusted workflow.
Failure mechanism: A malicious or compromised agent injects instructions, alters shared context, or exploits delegation logic so the receiving agent performs actions it should not have authorised.
Impact: Organisations can see data leakage, unsafe tool use, hidden lateral movement across agents, and cascading compromise when one agent’s trust is reused by others.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
OWASP Agentic AI Top 10 and MITRE ATLAS address the attack and risk surface, while NIST SP 800-53 Rev 5 sets the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| OWASP Agentic AI Top 10 | ASI03 — Identity & Privilege Abuse | Cross-agent attacks abuse delegated authority and agent trust boundaries. |
| ASI07 — Insecure Inter-Agent Communication | The term centers on harmful behaviour across agent communication channels. | |
| ASI08 — Cascading Failures | One compromised agent can trigger downstream failures across other agents. | |
| Recommendation — Constrain agent delegation and verify every privileged cross-agent action before execution. Authenticate inter-agent messages and reject untrusted or unauthenticated handoffs. Limit shared dependencies so one agent compromise cannot cascade through the system. | ||
| MITRE ATLAS | AML.T0118 — Autonomous AI Agent Communication | Cross-agent attacks use shared artefacts and direct channels between autonomous agents. |
| Recommendation — Map agent communication paths and monitor them for covert coordination or abuse. | ||
| NIST SP 800-53 Rev 5 | AC-6 — Least Privilege | Limiting agent permissions reduces what a compromised agent can trigger in peers. |
| Recommendation — Apply least privilege to every agent role and delegated tool path. | ||
Practitioner Guidance
Why practitioners should care: Cross-agent attack is a governance problem as much as a technical one, because it exposes where one agent is allowed to act for another. The practical question is whether each handoff, shared channel, and delegated capability has a clear owner and an enforceable boundary.
Common misunderstanding: Teams often secure the individual agent but leave the inter-agent relationship undercontrolled. That leaves the system vulnerable even when each agent appears well configured in isolation.
Practitioner takeaway: Treat every agent-to-agent path as a privileged interface, and design it so that trust is explicit, narrow, and revocable.
Related resources from NHI Mgmt Group
Deepen Your Knowledge
Free weekly newsletter
Subscribe to the NHI & AI Identity Journal
The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.
Bonus 33% off our NHI Course when you subscribe.
Reviewed and updated by the NHIMG editorial team on November 16, 2026.
NHI Mgmt Group — the #1 independent authority on Non-Human Identity, IAM, and Agentic AI security. nhimg.org