Use layered controls rather than a single approval step. Filter hostile input before the model runs, validate outputs before any side effect, and reserve human approval for high-impact actions such as spending, deleting, or policy changes. That combination reduces both prompt-injection risk and accidental tool misuse.
Why layered guardrails work better than a single approval step
An AI agent becomes safer when no single stage can directly trigger a harmful action. Input filtering reduces the chance that hostile instructions reach the model, output validation blocks unsafe commands before they execute, and human approval is reserved for actions with real blast radius. That separation matters because unsafe behaviour can come from deception, model error, or overbroad tool access.
Layering also prevents a common failure mode: teams put trust in one “final check” and assume the rest of the system is safe. In practice, the most dangerous actions are often assembled across steps, where a benign-looking prompt, a weak tool policy, and an over-permissive action path combine into one unsafe result.
When the action itself is reversible or low impact, automated checks should do most of the work. When the action changes money, data, access, or policy, the control objective shifts from convenience to containment. A good design treats approval as one control in a chain, not as a substitute for input hygiene or action-level policy.
Where unsafe actions usually enter the workflow
Unsafe agent behaviour usually starts before the model makes a decision. Prompt injection, poisoned context, and hostile instructions can steer the agent toward a misleading goal, especially when the agent reads untrusted content or mixes user input with operational instructions. The safer pattern is to separate untrusted text from control instructions and to inspect it for action-shaping content before it reaches the model.
The second entry point is the tool boundary. Even a well-behaved model can cause harm if it is allowed to call powerful tools too freely, reuse ambient credentials, or execute actions without per-request checks. Stronger designs externalise authorisation so the system evaluates each tool call against policy instead of assuming the model will self-limit.
The third entry point is the effect layer, where the agent actually changes state. Output validation matters here because the model may produce a plausible-looking command that is syntactically valid but operationally unsafe. That is where a policy engine, schema check, allow list, or action classifier should stop the request before any side effect occurs.
How to decide when human approval is needed
The right approval threshold is based on impact, not on how “smart” the model appears. Human review is most valuable when the action is hard to reverse, affects external parties, or can create privilege, financial, or policy changes. For routine actions, forcing review too early often creates noise, delays, and reviewer fatigue that weakens the whole control chain.
Approval is also most effective when it is attached to a specific decision, not a vague workflow milestone. The reviewer should know exactly what will happen, which system will be touched, and what the agent has already checked. That context makes it possible to approve the action itself, rather than merely rubber-stamp the agent’s summary.
For teams building this into production, the practical question is whether the workflow can safely auto-execute after validation. If the answer is no, the design should make the reviewer the final authority for that specific action class, while still keeping the upstream filtering and validation layers in place.
Risk and Threat Considerations
Unsafe agent actions create both accidental and adversarial exposure. A prompt-injection path can redirect the agent toward a harmful tool call, while overbroad permissions can turn a small instruction error into data loss, unwanted spending, or access changes. The main risk is not just that an agent makes a bad decision, but that it is able to carry that decision out with real authority.
Failure mechanism: The control fails when the agent treats untrusted input as actionable instruction, or when the tool layer allows execution without independent policy checks. A single approval gate is weak if hostile content can shape the request before review or if the review happens after the side effect is already queued.
Impact: The result can be destructive actions, privilege misuse, unauthorised state changes, or expensive mistakes that appear legitimate because they were made through an approved agent workflow. In high-autonomy systems, the blast radius grows quickly when actions are repeatable, cross-system, or difficult to roll back.
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 addresses the attack and risk surface, while NIST AI RMF, NIST SP 800-53 Rev 5 and NIST Zero Trust (SP 800-207) set the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| OWASP Agentic AI Top 10 | ASI02 — Tool Misuse | Agent tool calls can trigger unsafe actions if not constrained. |
| ASI03 — Identity & Privilege Abuse | Unsafe actions often come from excess authority, not model error alone. | |
| ASI01 — Agent Goal Hijack | Prompt injection can redirect an agent toward harmful objectives. | |
| Recommendation — Enforce per-action checks before any tool call that can change state. Limit agent privileges to the minimum needed for each task. Filter hostile instructions before they can alter the agent's goal. | ||
| NIST AI RMF | Govern | The topic is about AI risk governance and control layering. |
| Recommendation — Define approval thresholds and accountability for high-impact agent actions. | ||
| NIST SP 800-53 Rev 5 | AC-6 — Least Privilege | Reducing unsafe actions depends on constraining what the agent can do. |
| AU-2 — Event Logging | Agent actions need traceability when a tool or approval path goes wrong. | |
| Recommendation — Restrict each agent capability to the minimum permission set required. Log agent decisions, tool calls, and high-impact approvals for review. | ||
| NIST Zero Trust (SP 800-207) | Never Trust, Always Verify | The answer relies on verifying each request and action, not trusting the agent. |
| Recommendation — Verify every action request before permitting execution. | ||
Practitioner Guidance
What to prioritise: Put the strongest controls on the highest-impact actions first. If an agent can spend money, delete records, change permissions, or publish externally, treat those as separate action classes with stricter validation than ordinary read-only or drafting tasks.
What to verify: Confirm that the system checks intent, output, and side effect separately. A good implementation proves that the prompt filter, action validator, and human approval are independent barriers, not three labels for the same decision point.
Decision rule: If the action can create material external impact, require human approval after automated validation, not before it. If the action is low-risk and fully reversible, keep the path automated but still log the decision and the inputs that shaped it.
Practitioner takeaway: The safest AI agents are not the ones that “think carefully”, but the ones that can only act within narrow, verifiable bounds, with human judgment reserved for the few actions that truly deserve it.
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 4, 2026.
NHI Mgmt Group — the #1 independent authority on Non-Human Identity, IAM, and Agentic AI security. nhimg.org