Unbounded consumption is the failure to cap how much compute, token volume, or downstream cost an AI request can consume. In LLM environments, it turns legitimate use into a financial and availability attack surface when input size, retries, and model complexity are left unchecked.
Expanded Definition
Unbounded consumption describes a design and governance failure where an AI or automation workflow can absorb unlimited compute, tokens, retries, tool calls, or downstream service costs. In NHI and agentic AI environments, the term matters because the identity that authorises the action may be legitimate while the economic impact remains unconstrained.
This is distinct from ordinary cost overruns. Unbounded consumption usually arises when a prompt, agent, or API workflow has no hard limits on context size, recursion depth, model selection, or external calls. In practice, that makes the request path itself a control point, not just an application performance concern. Guidance from NIST Cybersecurity Framework 2.0 reinforces that resilience must include resource governance, while NHI management requires visibility into how identities trigger spending and execution. NHI Mgmt Group notes that 97% of NHIs carry excessive privileges, which makes cost-amplifying workflows especially difficult to contain when access and usage are not tightly scoped. The most common misapplication is treating token and tool usage as an application tuning issue, which occurs when organisations fail to set enforcement thresholds at the identity, model, or gateway layer.
Examples and Use Cases
Implementing controls against unbounded consumption often introduces friction, because strict limits can reduce flexibility for legitimate high-volume tasks and require careful exception handling.
- An agent is allowed to call a high-cost model repeatedly until a task succeeds, so a malformed prompt can create a runaway billing event.
- A customer support copilot can expand conversation history without a token cap, causing context growth that degrades latency and inflates inference costs.
- A tool-using workflow can loop through the same retrieval and summarisation steps without a retry ceiling, exhausting both API quotas and compute budgets.
- A service account used by an internal AI assistant can trigger bulk document processing, and a missing per-identity quota allows one workload to consume shared capacity.
The operational pattern is closely related to identity governance failures described in Ultimate Guide to NHIs, where over-permissioned and poorly observed non-human identities become difficult to constrain. Implementation teams also borrow from NIST Cybersecurity Framework 2.0 to align rate limiting, monitoring, and response with broader operational risk objectives. In mature environments, unbounded consumption controls are paired with per-agent quotas, model tier restrictions, and hard stop conditions before requests hit expensive back-end systems.
Why It Matters in NHI Security
Unbounded consumption is a security issue because it converts authorised access into a denial-of-wallet, service degradation, or cascading failure scenario. In agentic systems, the same identity that can legitimately invoke a tool can also drive unexpected scale if usage is not bounded by policy. That is especially dangerous when secrets, service accounts, and API keys are embedded in automation paths that already lack strong lifecycle governance.
NHI Mgmt Group reports that only 5.7% of organisations have full visibility into their service accounts, which means most teams cannot reliably see which identities are generating cost or load. Without that visibility, finance, operations, and security may all miss the same warning signs until limits are breached. The control problem is not just technical throttling; it also includes approval paths, anomaly detection, and per-identity entitlements that keep AI execution proportional to intent. The 80% of identity breaches involving compromised non-human identities highlighted in the Ultimate Guide to NHIs underscore how quickly legitimate automation can become an abuse channel when governance is weak. Organisations typically encounter the true impact only after billing spikes, quota exhaustion, or service outages, at which point unbounded consumption becomes operationally unavoidable to address.
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 OWASP Non-Human Identity Top 10 address the attack and risk surface, while NIST CSF 2.0, NIST Zero Trust (SP 800-207) and NIST AI RMF set the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| OWASP Agentic AI Top 10 | Addresses agent runaway loops and excessive tool/model usage that drive unbounded consumption. | |
| OWASP Non-Human Identity Top 10 | NHI-04 | Maps to governance failures where NHI-driven automation lacks usage limits and monitoring. |
| NIST CSF 2.0 | PR.PS-1 | Supports platform protection and resource controls for AI workloads and service identities. |
| NIST Zero Trust (SP 800-207) | Zero Trust requires continuous verification and least privilege for every AI action path. | |
| NIST AI RMF | Risk management for AI includes resource abuse, cost volatility, and operational resilience. |
Set hard caps on agent steps, tool calls, retries, and model escalation before deployment.
Related resources from NHI Mgmt Group
Deepen Your Knowledge
Reviewed and updated by the NHIMG editorial team on August 28, 2026.
NHI Mgmt Group — the #1 independent authority on Non-Human Identity, IAM, and Agentic AI security. nhimg.org
