Security teams should test the full AI stack as one attack surface, not as isolated components. Use probes that cover prompt injection, agent goal hijacking, tool poisoning, privilege escalation, and secret exposure across LLMs, autonomous agents, and MCP servers. The goal is to find chained failures before an attacker can move from model compromise to unauthorized actions.
How to structure red teaming across the stack
Red team the system as a single attack path, not as three disconnected tests. The model layer, orchestration layer and MCP layer should be exercised together so you can see whether a weakness in one layer can be chained into privileged tool use, data exposure or unsafe actions in another. That is the difference between finding a bug and finding a real compromise path.
At the model layer, the focus is on whether the system can be manipulated into unsafe instructions, deceptive delegation or policy bypass. At the orchestration layer, the question is whether the agent can be induced to choose the wrong goal, call the wrong tool, or escalate beyond its intended scope. At the MCP layer, the question becomes whether the server trusts the wrong client, forwards tokens incorrectly, or exposes capabilities that the agent should never inherit.
Good testing makes those boundaries visible. A red team exercise should prove whether a prompt injection can alter tool selection, whether a poisoned tool or resource can change downstream behaviour, and whether a valid session can be turned into an action that the original user never authorised. NHIMG’s Agentic AI Security Guide is useful here because it frames the layered attack surface in terms of inputs, tools, orchestration and identity rather than treating agent failures as isolated events.
What to probe in the model, orchestration and MCP layers
Start with the model because it often supplies the first foothold. Test prompt injection, indirect prompt injection, hidden instruction smuggling and content that tries to reframe policy or task priority. Then move up one layer and see whether the orchestration logic preserves the right objective when the model output is ambiguous, adversarial or overloaded with irrelevant context.
The orchestration layer is where many agentic failures become material. Probe goal hijacking, unsafe tool chaining, overbroad retries, false confidence in a low-quality model response and missing policy checks before action execution. If the agent can still reach a sensitive function after the original intent has been distorted, the control plane is too trusting. Top 10 Agentic AI Identity Issues maps these failure modes to privilege abuse, overprivileged agents and approval bypass, which is exactly where a red team should push.
For MCP, test the contract between client, server and upstream tools. Look for token forwarding that crosses trust boundaries, audience confusion, capability drift between servers, and assumptions that every request arriving through MCP is inherently legitimate. The practical question is simple: can the agent or a maliciously shaped tool response cause the server to do something that should have required a different principal, a narrower token or a separate approval step? Model Context Protocol: Authorization specification is the right reference point for checking whether the implementation preserves those boundaries.
What good evidence looks like in a red team result
A useful finding is not just “the model was fooled”. It should show a concrete chain: the initial input, the point at which the agent’s intent changed, the tool or server that accepted the wrong request, and the final action that should not have been possible. That chain matters because remediation differs depending on where the failure actually occurred.
If the break happened in the model, the fix may be better input handling, stronger instruction separation or refusal logic. If it happened in orchestration, the fix may be stricter pre-action policy checks, tighter tool routing or step-up approval before sensitive actions. If it happened in MCP, the fix may be stronger server-side authorization, token audience validation or removal of passthrough trust assumptions. The same visible symptom can have very different root causes, so the red team report should preserve the sequence, not just the outcome.
Teams should also capture whether the compromised path was reusable. A one-off prompt trick is serious, but a path that consistently reaches the same tool or server is more important because it can be operationalised by an attacker. That is where exploitability becomes governance, not just a lab result.
Risk and Threat Considerations
Agentic systems fail in chains, and that makes layered red teaming essential. A model-level compromise may not matter by itself, but if it can steer orchestration into privileged tool use or abuse MCP trust boundaries, the result can be unauthorized actions, data exfiltration or lateral movement through connected services.
Failure mechanism: An attacker or malicious input manipulates the model, influences orchestration decisions, or abuses MCP trust and authorization assumptions so the agent performs actions outside its intended scope.
Impact: The organisation may lose control over tool execution, data access, approvals and downstream side effects, especially when the same session or token can be reused across multiple steps.
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 ATT&CK address the attack and risk surface, while NIST AI RMF sets the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| OWASP Agentic AI Top 10 | ASI01 — Agent Goal Hijack | Goal hijack is central to cross-layer agent red teaming. |
| ASI02 — Tool Misuse | The question focuses on unsafe tool use across orchestration and MCP. | |
| ASI03 — Identity & Privilege Abuse | Privilege escalation and unauthorized actions are explicit red-team targets. | |
| Recommendation — Test whether adversarial inputs can redirect agent goals before tool execution. Validate that tools cannot be called outside their intended purpose or scope. Check whether the agent can act with privileges beyond the current request. | ||
| MITRE ATT&CK | T1059 — Command and Scripting Interpreter | Agentic tool use often turns malicious instructions into executable commands. |
| T1110 — Brute Force | Credential and access abuse can follow chained failures in agentic systems. | |
| T1552 — Unsecured Credentials | Secret exposure is an explicit red-team objective in the question. | |
| Recommendation — Map tool execution paths that can be abused to run attacker-influenced commands. Hunt for repeated auth attempts and weak rate-limiting around agent access paths. Search for places where prompts, logs or tools expose usable secrets. | ||
| NIST AI RMF | GV.1 — Govern AI Risk | Agentic red teaming is a risk-governance activity across the AI stack. |
| Recommendation — Define and own the AI risk posture for model, orchestration and tool layers. | ||
Practitioner Guidance
What to prioritise: Red team the highest-consequence action first, not the prettiest demo path. If the agent can touch production data, external systems or identity-sensitive tools, test those chains before spending time on low-impact prompt failures.
What to verify: Every finding should identify which layer failed and whether the failure was caused by model behaviour, orchestration logic or MCP trust handling. If you cannot localise the break, you do not yet have a remediation plan.
Decision rule: If a prompt injection can only distort text, treat it as a model weakness; if it can change tool selection or server-side action, treat it as a control failure across the stack and escalate immediately.
Practitioner takeaway: The test is not whether each layer looks secure on its own, it is whether a bad instruction can survive translation from model output to orchestrated action to MCP request.
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 3, 2026.
NHI Mgmt Group — the #1 independent authority on Non-Human Identity, IAM, and Agentic AI security. nhimg.org