Adversarial Investigation Resource Exhaustion (AIRE): Defensive Workload Amplification in AI-Enabled Security Operations
C/O Futures AI Initiatives Research Report
Kaden K. Bunker, Jean-michel Newberg, and Robert J. Bunker, SME Contributors
Human-AI Teaming Product
November 2026

When Correct Investigation Becomes the Attack Surface
AI-enabled security systems do not necessarily have to be deceived to be weakened. In some cases, they may only have to be kept busy.
Security operations are increasingly delegating alert triage, evidence acquisition, correlation, diagnosis, threat-intelligence enrichment, and response planning to large-language-model and agentic systems. This expansion of defensive autonomy introduces a resource-management problem that goes beyond whether an AI system reaches the correct conclusion.
An adversary may be able to influence the security environment in ways that cause an AI-enabled defender to perform legitimate but disproportionately expensive investigative work. The defensive system may correctly conclude that an event is benign while still consuming model inference, context, tool queries, API capacity, cloud resources, investigation workers, queue capacity, financial budget, and human escalation time along the way.
C/O Futures proposes Adversarial Investigation Resource Exhaustion (AIRE) as a framework for examining this problem.
Rather than treating token consumption or model inference as the only resource at risk, AIRE treats the defensive investigation itself as an adversarially consumable resource. The central question becomes not simply whether an AI defender is correct, but how much defensive investigation a security-relevant event can induce relative to the actual threat it represents.
Key Findings
Cross-source investigation fan-out emerged as the strongest isolated workload mechanism tested. Broader evidence requirements consistently increased tool use and simulated evidence context across matched security scenarios, while ambiguity and recurrence alone did not produce the same consistent effect.
Broader evidentiary requirements produced progressively greater defensive work until bounded investigation policies reached their resource ceiling.
Once that ceiling was reached, the failure mode shifted from additional expenditure to incomplete investigation, creating a resource/completeness tradeoff.
Memory and deduplication substantially reduced modeled recurrence workload by limiting unnecessary reacquisition of previously investigated evidence.
Increased per-alert investigative demand created greater queueing pressure for genuine incidents even when total alert volume remained unchanged.
Reserving some defensive capacity for genuine incidents reduced their modeled delay under high-fan-out workload, but shifted congestion onto the remaining unreserved investigation capacity.
The study does not establish a production exploit against a commercial AI security operations center and does not claim measured token, dollar, GPU, energy, or analyst-cost amplification.
The findings support a narrower systems-security conclusion: investigation breadth, resource controls, memory, stopping policy, and shared capacity should be treated as part of the attack surface of increasingly autonomous defensive systems.
What Is Adversarial Investigation Resource Exhaustion?
C/O Futures defines Adversarial Investigation Resource Exhaustion as:
The deliberate generation or manipulation of security-relevant activity or information in order to cause an AI-enabled defensive system to expend disproportionate computational, financial, operational, or human resources conducting otherwise legitimate or apparently legitimate investigations.
The distinction is important.
AIRE is not primarily about directly sending an AI model an enormous prompt, forcing recursive output, or flooding an inference endpoint. Those behaviors are already addressed by existing work on unbounded consumption, Denial-of-Wallet, model-level denial of service, reasoning-consumption attacks, and agentic tool amplification.
Instead, AIRE begins in the security environment being observed.
An adversary influences alerts, logs, entities, correlations, telemetry, or other security-relevant information. The defensive system then decides—legitimately—to investigate that activity.
The resulting investigation may expand across endpoint data, identity systems, network telemetry, threat intelligence, historical cases, external services, and multiple agentic tools.
The defender does not have to be wrong.
The adversarial value may lie entirely in the amount of work required to become confident that the event is benign.
Beyond Denial-of-Wallet and Alert Flooding
The broader idea that attackers can force AI systems to consume resources is not new.
Prior research has examined Denial-of-Wallet attacks against pay-per-use systems, adversarial examples designed to increase neural-network latency and energy consumption, prompts that extend model generation, attacks that induce excessive reasoning, retrieval-based inference-cost attacks, and agentic tool chains that expand the number of actions required to complete a task.
Security research has also long recognized the problem of defensive overload. Informational Denial-of-Service describes how adversarial feints can consume limited human attention and make genuine activity more difficult to identify.
AIRE sits at the intersection of these existing ideas but focuses on a narrower problem:
What happens when adversary-influenced security activity causes an AI-enabled defender to perform legitimate investigation work across multiple defensive systems, and the attack succeeds through the cost or capacity consequences of that investigation rather than through an incorrect conclusion?
That distinction becomes increasingly relevant as AI security systems evolve from assisting human analysts with summarization toward conducting more autonomous investigation.
The Investigation Graph Matters More Than the Token Count
The AIRE project began with a relatively token-centric intuition: an attacker might force an AI defender to spend excessive tokens investigating fabricated or low-value activity.
The research ultimately suggests that this framing is too narrow.
Token consumption is only one implementation-specific resource. Modern AI-enabled investigations can involve model inference, retrieved context, repeated evidence acquisition, external tools, API calls, parallel agents, retries, workflow actions, concurrency, queue occupancy, and human escalation.
The more useful unit of analysis is therefore the investigation graph.
An investigation graph consists of the evidence branches, tools, correlations, retries, memory accesses, parallel activities, and escalation steps created when a defensive system attempts to resolve a security event.
Several architectural properties become important:
Depth describes the number of sequential investigative steps.
Breadth or fan-out describes how many distinct evidence domains or parallel investigative branches must be consulted.
Recurrence captures repeated or reopened investigative work.
Memory and reuse determine whether prior investigative state can prevent unnecessary repetition.
Budget determines how much resource consumption is permitted for a case.
Concurrency determines how much investigative work can execute simultaneously.
Priority reservation determines whether some capacity remains available for higher-value defensive work.
Stopping criteria determine when sufficient evidence has been gathered to end an investigation.
AIRE risk therefore emerges from the interaction between the security event, the structure of the resulting investigation, the policies governing that investigation, and the resources available to the defensive system.
How the AIRE Study Was Designed
The C/O Futures study was deliberately constructed as a cost-free, controlled, synthetic experiment intended to isolate investigation structure rather than reproduce vendor-specific AI economics.
All security scenarios were fictional. No live system was scanned, exploited, flooded, or contacted. The test environment contained no malware execution, credential attacks, exploit payloads, or instructions for producing harmful network traffic.
The synthetic defensive environment exposed several common categories of security
evidence, including endpoint context, identity context, network context, threat-intelligence context, and prior case history.
The exploratory study compared several matched conditions: routine benign activity, natural false positives, genuine incidents, ambiguity-focused AIRE cases, fan-out-focused AIRE cases, recurrence-focused AIRE cases, and mixed AIRE cases.
A second controlled study then progressively increased the breadth of evidence required to resolve otherwise comparable benign events.
The resulting investigation ladder was evaluated under three deterministic defensive policies.
A minimalist policy used a strict investigation budget and stopped early when resolving evidence became available.
A bounded policy allowed broader investigation while retaining a finite resource ceiling.
An exhaustive policy consulted every available evidence domain even when sufficient resolving evidence had already appeared.
The study also modeled recurrence under different memory policies and placed investigation demand into a finite shared-capacity queue model to examine how expensive investigations could influence genuine defensive work.

Fan-Out Emerged as the Strongest Isolated Workload Driver
The exploratory pilot produced one of the report's most important findings.
Against matched natural false positives, fan-out-focused AIRE cases consistently required broader tool use and greater simulated evidence context across the evaluated security families.
Ambiguity alone did not produce the same result.
The ambiguity-focused condition generated little change in tool use or overall evidence volume. Recurrence alone likewise did not increase tool use under the study's case-history behavior.
Mixed AIRE cases increased workload, but their behavior was less uniform than isolated fan-out.
This result challenged the project's original assumption that ambiguity itself would reliably create expensive investigation.
Instead, ambiguity appears to matter primarily when it forces the defender to acquire additional evidence.
The security implication is significant.
A low-risk event does not necessarily have to produce hundreds or thousands of alerts to consume substantial defensive capacity. A relatively small number of events that each require investigation across many independent evidence sources could impose meaningful resource demand without resembling a conventional volumetric flood.
Work Grows Until the Resource Ceiling
The controlled fan-out experiment reinforced this result.
As the evidence required to resolve a case became progressively broader, defensive investigation expanded accordingly.
Narrow cases were resolved quickly.
Moderate cases required additional evidence.
Broad cases required several evidence domains.
Very broad cases approached the bounded policy's investigation ceiling while still reaching resolution.
Eventually, however, the evidentiary requirement exceeded the resources the bounded policy permitted.
At that point, the defensive system stopped spending—but it also stopped before reaching the evidence necessary to resolve the case.
C/O Futures describes this resource/completeness transfer as safeguard inversion.
The resource control works exactly as intended: it prevents unlimited investigation expenditure.
But the same ceiling becomes a threshold beyond which defensive completeness degrades.
This exposes a fundamental design tension for autonomous investigation.
Aggressive stopping reduces the attack surface for resource amplification but increases the possibility of unresolved cases.
Aggressive evidence gathering increases investigative completeness but creates a larger resource-exhaustion surface.

Correctness Alone Is Not Enough
AI-security evaluations frequently focus on whether a model or agent reaches the correct final answer.
AIRE demonstrates why that metric is incomplete for autonomous defense.
An AI investigator might correctly classify an event as benign and still have performed an adversarially expensive amount of work to reach that conclusion.
This means defensive AI should be evaluated not only on disposition accuracy but also on process efficiency under adversarially selected workloads.
Relevant measurements may include evidence breadth, tool use, context growth, retries, queue occupancy, investigation time, resource-control activation, and downstream human escalation.
A system that remains accurate while consuming uncontrolled shared resources may still be operationally vulnerable.
Memory Can Reduce Repeated Work
The study's recurrence model showed that repeated investigation becomes most expensive when defensive architecture lacks reusable state.
Without memory, recurring activity can cause the system to repeatedly rebuild the same investigation graph and reacquire evidence it has already seen.
Case-summary memory reduces this demand by allowing prior investigative work to inform subsequent events.
Deduplication and early stopping reduce it further by recognizing when recurring activity does not warrant rebuilding the full investigation.
The result changes how recurrence should be understood.
Repetition is not inherently expensive.
It becomes an amplification path when the architecture repeatedly reacquires information that could safely be reused, summarized, cached, or suppressed.
Memory is therefore a potentially important AIRE mitigation—but it also creates a new security dependency.
Stale, incorrect, or attacker-influenced historical conclusions could cause necessary reinvestigation to be suppressed. Defensive architects must therefore balance resource
reuse against evidence freshness, provenance, and integrity.
Expensive Investigations Can Delay Genuine Incidents Without More Alerts
The finite-capacity simulation connects per-investigation resource use to a broader defensive consequence.
Rather than increasing total alert volume, the model replaced part of ordinary investigative workload with more demanding high-fan-out investigations.
The total number of alerts remained fixed.
The amount of work required to process them did not.
As the proportion of high-fan-out investigations increased, more service demand was placed on the same shared defensive capacity, producing greater queueing pressure for genuine incidents.
This demonstrates that alert count alone is not sufficient to understand defensive workload.
Per-alert investigative demand is an independent capacity variable.
A conventional alert flood can still create substantial delay, and the study does not claim that fan-out is universally more damaging than volumetric overload. The relative impact changes depending on available concurrency and other architectural assumptions.
The important finding is narrower:
An adversary may be able to increase defensive delay by increasing the amount of legitimate investigation required per event—even without increasing the number of alerts.
Protected Capacity Helps, but the Congestion Moves
The study also examined whether some shared investigative capacity could be protected for genuine incidents.
In the modeled environment, reserving capacity for higher-priority work reduced genuine-incident queueing delay during high-fan-out workloads.
The mitigation was not free.
Ordinary investigations were forced to compete for a smaller unreserved resource pool and therefore experienced greater congestion.
The broader principle is that low-value or unresolved investigative work should not be allowed to consume all capacity required for higher-priority defensive activity.
This does not necessarily require a literal dedicated worker or agent. Protected capacity could be implemented through workload classes, queues, scheduling priorities, reserved concurrency, or other architectural mechanisms.
Implications for Government, Critical Infrastructure, and National Security
AIRE is relevant beyond commercial security operations centers.
Government and national-security organizations are increasingly considering AI for cyber defense, intelligence triage, fraud detection, infrastructure monitoring, operational decision support, and other workflows in which an adversary may intentionally shape the information environment presented to the system.
Different sectors will also make different resource-allocation decisions.
A critical infrastructure operator primarily concerned with system availability may tolerate continued investigation or even continued activity against a direct target if interrupting the system would create a greater operational risk.
A defense environment with a stronger confidentiality requirement may accept greater availability disruption in exchange for rapidly containing suspicious activity.
The policy implication is not that high-consequence organizations should avoid AI autonomy.
It is that autonomy expands the attack surface from decision correctness into decision-process economics and capacity management.
Adversarial testing should therefore include workloads that are low in underlying threat but high in legitimate investigative demand and should explicitly examine what happens when resource budgets, queues, concurrency limits, or human escalation channels become saturated.
Defensive Design Principles for AIRE Resilience
The report identifies several principles for reducing exposure to adversarial investigation resource exhaustion.
Investigation admission control can prevent every alert from automatically receiving unrestricted agentic investigation. Lightweight deterministic filtering, severity rules, deduplication, and historical context can determine which events justify more expensive analysis.
Per-investigation resource budgets should track consumption at the investigation or agent-run level rather than only per API key or individual model call. Resource accounting should include tool use and downstream services—not just tokens.
Fan-out controls should make the number of evidence domains, subagents, and tool branches activated by a case observable and attributable.
Memory and deduplication can safely reuse prior evidence and prevent recurring activity from automatically rebuilding an entire investigation graph.
Cost-aware stopping criteria should consider evidentiary sufficiency and the marginal value of additional investigation. Agents should not query additional systems merely because those tools are available.
Priority reservation can protect some shared capacity for unresolved or higher-priority threats.
Graceful security degradation should define what happens when a resource ceiling is reached. Rather than silently terminating, a system could move to deterministic fallback analysis, delayed enrichment, a smaller model, a constrained toolset, or explicit human escalation based on risk.
Finally, resource telemetry itself can become a defensive signal. Sudden increases in fan-out, repeated evidence acquisition, retries, context growth, or queue occupancy may indicate either architectural inefficiency or adversarial workload shaping.
What the Study Does—and Does Not—Establish
The AIRE report intentionally maintains a narrow claim boundary.
It proposes AIRE as a framework for treating defensive investigation capacity as an adversarially consumable resource.
The controlled study found that fan-out consistently increased investigative tool use and simulated evidence workload relative to matched ordinary false positives.
The fan-out ladder demonstrated a resource/completeness tradeoff under bounded investigation.
Modeled memory reduced recurrence workload.
The finite-capacity simulation showed that greater per-alert service demand can delay genuine incidents while alert volume remains fixed.
The study does not, however, establish that AIRE has been successfully executed against a commercial AI SOC platform.
It does not provide a specific real-world response slowdown.
It does not claim that the project discovered resource exhaustion, alert flooding, agent fan-out, telemetry manipulation, or the other underlying concepts on which the framework builds.
It also does not claim measured provider-token amplification, dollar-cost amplification,
GPU utilization, energy consumption, or analyst labor effects.
The report should therefore be understood as a proof-of-concept systems-security
framework and controlled methodology, not as evidence of a production vulnerability in a named commercial platform.
The Next Research Step
The most important next step is external validation.
A future study could instrument a real or open-source agentic investigation platform and safely replay synthetic security events while recording actual model and tool telemetry.
Such testing could begin to measure how the mechanisms identified in the synthetic study translate into real model calls, investigation latency, tool use, concurrency demand, queue occupancy, financial consumption, and human escalation.
Future work should also examine how adaptive models behave under fan-out pressure, how memory integrity affects the resource/completeness tradeoff, how different security sectors should allocate protected capacity, and what attacker effort would actually be required to create high-fan-out investigative conditions.
AIRE Changes the Question
The central lesson of the research is not simply that artificial intelligence can consume too many tokens.
As defensive systems become more autonomous, the investigation process itself becomes part of the security architecture—and therefore part of the attack surface.
The important question is:
How much defensive investigation can a security-relevant event induce relative to its underlying threat?
A resilient autonomous defender must not only reach the correct conclusion. It must reach that conclusion while preserving enough computational, operational, financial, and human capacity to continue investigating the threats that actually matter.
Download the Full C/O Futures AI Initiatives Research Report
The complete 21-page C/O Futures AI Initiatives Research Report contains the AIRE threat framework, related-work analysis, controlled synthetic methodology, research questions, exploratory and fan-out studies, recurrence and finite-capacity models, defensive design principles, limitations, ethics and reproducibility discussion, claim boundaries, references, and supporting analysis.
Adversarial Investigation Resource Exhaustion (AIRE): Defensive Workload
Amplification in AI-Enabled Security Operations
Kaden K. Bunker, Jean-michel Newberg, and Robert J. Bunker, SME Contributors
Related C/O Futures Research
The original C/O Futures AIRE op-ed introduced the concept of adversarial resource exhaustion against agentic AI defenses and established the initial research direction that developed into the broader AIRE framework and controlled study presented in this report.
Related C/O Futures research has examined how generative AI capabilities can affect security operations and adversarial activity, reinforcing the need to evaluate emerging AI systems not only for capability but also for their changing threat surfaces.
C/O Futures has also examined the evolving intersection of emerging technologies, adversarial behavior, extremist activity, and the security implications created by increasingly technology-dependent environments.
Work With C/O Futures
C/O Futures provides specialized research, threat analysis, intelligence support, briefings, red teaming, and training related to artificial intelligence, emerging threats, adversarial technologies, and evolving operational environments.
Follow C/O Futures Research
Receive newly published C/O Futures research notes and analyses by email.

Comments