Serverless Under Attack: Simulating Cloud Security Risks

by

Author: Denis Avetisyan

Researchers have developed a new simulator to proactively analyze the security vulnerabilities inherent in serverless computing architectures.

Kumo’s architecture establishes a foundation for adaptable and scalable distributed computing through a modular design.
Kumo’s architecture establishes a foundation for adaptable and scalable distributed computing through a modular design.

Kumo enables detailed investigation of DoS attacks, co-location risks, and resource contention in serverless platforms, distinguishing scheduler-driven isolation from general resource exhaustion.

While serverless computing simplifies application deployment, it simultaneously obscures system-level behaviors that introduce potential security vulnerabilities. To address this, we present ‘Kumo: A Security-Focused Serverless Cloud Simulator’, a discrete-event simulator designed for controlled analysis of risks arising from scheduling and resource sharing in serverless platforms. Our results demonstrate that scheduler choice significantly impacts co-location attack surfaces, while denial-of-service behavior is largely governed by system-level factors like queuing policy and cluster capacity. How can a deeper understanding of these interactions inform the design of more secure and resilient serverless architectures?

The Paradox of Efficiency: Serverless Security in a Shared Landscape

Serverless computing, while lauded for its potential to dramatically scale applications and reduce operational costs, fundamentally alters the security landscape through its reliance on extensive resource sharing. This architectural shift means multiple applications, potentially belonging to different users or organizations, can execute within the same underlying infrastructure. While this maximizes resource utilization and minimizes expenses, it also introduces novel attack vectors; a compromised function could potentially access or interfere with others sharing the same environment. Traditional security models, designed for static infrastructure with clear boundaries, struggle to adapt to this dynamic, multi-tenant reality, creating a paradox where the pursuit of efficiency inadvertently increases security risks and demands innovative protective measures.

A fundamental challenge within serverless architecture centers on the trade-off between performance and security, specifically concerning container reuse. To avoid the performance bottleneck of ‘Cold Starts’ – the delay when a function is invoked after inactivity – serverless platforms aggressively recycle containers to serve multiple requests. However, this practice introduces the risk of ‘malicious co-location’, where a compromised or malicious function is scheduled to run within the same container as a legitimate one. This proximity allows for potential data breaches, code injection, or resource hijacking, as the functions share the same memory space and kernel. Effectively, minimizing latency demands a level of resource sharing that simultaneously expands the attack surface, creating a complex security balancing act for developers and platform providers.

Conventional security methodologies, designed for static infrastructure and dedicated resources, struggle to address the ephemeral and shared characteristics of serverless computing. These systems typically rely on perimeter-based defenses and known entity identification, failing to account for the constantly shifting attack surface created by dynamic resource allocation and the inherent multi-tenancy. This disconnect leaves serverless applications susceptible to novel threats like function hijacking, data breaches stemming from noisy neighbor effects, and compromised dependencies introduced through shared runtime environments. Consequently, organizations adopting serverless architectures must fundamentally rethink their security posture, prioritizing fine-grained access control, robust isolation mechanisms, and continuous monitoring tailored to the unique challenges posed by this rapidly evolving paradigm.

Different scheduling algorithms impact the degree of co-location between attacker and victim functions in a large-scale serverless environment.
Different scheduling algorithms impact the degree of co-location between attacker and victim functions in a large-scale serverless environment.

Demonstrating the Attack Surface: Co-location and Denial-of-Service

Co-location attacks represent a substantial security risk, predicated on an attacker’s ability to schedule malicious code execution in close proximity to a targeted victim process. This proximity allows for interference with victim performance, potentially through cache contention or shared resource exhaustion, and creates opportunities for data theft via side-channel attacks or direct memory access. Simulation results indicate that the selection of scheduling algorithm significantly impacts the probability of successful co-location; certain schedulers exhibited co-location probabilities orders of magnitude higher than others under identical attack conditions, demonstrating a critical vulnerability dependent on system configuration.

Availability attacks and Denial-of-Service (DoS) exploits function by inducing resource contention, specifically targeting shared resources such as CPU cycles, memory bandwidth, or network capacity. Experimental results demonstrate a direct correlation between attacker intensity and the rate at which legitimate requests, or “victim” packets, are dropped. Increased attacker load consistently resulted in a measurable increase in victim drop rate, indicating that these attacks effectively degrade service quality and can ultimately render systems unavailable by overwhelming resources and preventing legitimate operations from completing. This behavior confirms that resource contention is a key mechanism by which these attacks achieve their disruptive effects.

The scheduler placement process is central to the feasibility of both co-location and denial-of-service attacks in shared computing environments. Specifically, the algorithm responsible for assigning functions to physical cores directly determines the proximity of potentially malicious and benign processes. Inadequate scheduling can result in high co-location probability, increasing the likelihood of interference or data exfiltration. Furthermore, the scheduler’s handling of resource allocation dictates the extent to which an attacker can induce resource contention and negatively impact victim performance or availability. Consequently, the development and deployment of secure scheduling algorithms – those incorporating mechanisms to minimize adverse co-location and mitigate resource contention – are essential for protecting shared infrastructure.

Increasing attacker intensity in Case Study B demonstrates a direct correlation with reduced victim availability and increased tail latency.
Increasing attacker intensity in Case Study B demonstrates a direct correlation with reduced victim availability and increased tail latency.

Kumo: A Platform for Proactive Serverless Security Analysis

Kumo is a simulation platform developed to address the unique security challenges present in serverless computing environments. Unlike traditional infrastructure, serverless architectures introduce complexities in resource allocation and function isolation, creating new attack surfaces. Kumo facilitates security analysis by providing a controllable and repeatable environment for modeling serverless deployments and simulating various attack vectors. The platform focuses specifically on the dynamics of function execution, resource contention, and potential vulnerabilities within these systems, enabling researchers and developers to proactively identify and mitigate risks before production deployment. Its design emphasizes the ability to test security measures in a realistic, yet isolated, serverless context.

Kumo employs Discrete-Event Simulation (DES) to model the ephemeral and stateful nature of serverless functions and their interactions with underlying resources. DES allows for the precise tracking of events-such as function invocations, resource allocations, and network communications-over time, providing a detailed representation of system behavior. This is coupled with realistic Workload Modeling, where function execution times, invocation rates, and data dependencies are derived from or parameterized by observed serverless application traces. The combination of DES and accurate workload models allows Kumo to replicate the dynamic scaling, concurrency, and resource contention characteristic of serverless environments, facilitating analysis beyond what is achievable with traditional monitoring or static analysis techniques.

Kumo facilitates the evaluation of serverless function scheduling algorithms under simulated attack conditions. Researchers can test the performance of algorithms including the Random Scheduler, DoubleDip Scheduler, and Helper Scheduler, analyzing their behavior when subjected to various security threats. Comparative analysis revealed that the Helper and OpenWhisk schedulers demonstrated a significant delay-several orders of magnitude-in function co-location compared to the Random Scheduler. This metric indicates a substantial difference in how quickly these algorithms place potentially vulnerable functions on the same execution environment, impacting the potential for lateral movement during an attack.

Kumo functions as a pre-deployment validation environment for security improvements in serverless architectures. This allows researchers and developers to proactively identify vulnerabilities and assess the effectiveness of proposed mitigations without impacting live systems. By simulating realistic attack vectors and workload conditions, Kumo enables quantitative evaluation of security enhancements, providing data-driven insights into their performance and impact on overall system resilience. This proactive testing reduces the risk of deploying flawed security measures and strengthens the ability of serverless applications to withstand real-world threats, ultimately improving the reliability and trustworthiness of the platform.

Towards Autonomous Resilience: The Future of Serverless Security

Recent investigations, notably those conducted by Kumo, highlight a critical shift in serverless security paradigms. Traditional approaches have largely relied on reactive defenses – responding to threats only after they manifest. However, serverless architectures, with their ephemeral and distributed nature, demand a move towards proactive security measures. Kumo’s work demonstrates the efficacy of identifying potential vulnerabilities before deployment through rigorous simulation and pre-emptive mitigation strategies. This involves modeling potential attack vectors and implementing safeguards during the development lifecycle, rather than solely relying on runtime protections. By anticipating and addressing weaknesses proactively, serverless systems can significantly enhance their resilience and minimize the impact of evolving cyber threats, fostering a more secure and dependable cloud environment.

Serverless systems, while offering scalability and cost efficiency, present unique security challenges demanding a shift towards proactive defense. Recent advancements highlight the efficacy of vulnerability identification and mitigation through rigorous simulation-a process enabling developers to anticipate potential attacks before deployment. This approach allows for the systematic testing of system responses to various threat vectors, including injection attacks, denial-of-service scenarios, and data breaches, within a controlled environment. By pinpointing weaknesses and implementing preventative measures during the development lifecycle, systems can be hardened against a broad spectrum of malicious activity. Ultimately, this simulation-driven methodology fosters the creation of resilient serverless architectures capable of maintaining operational integrity and protecting sensitive data, even under duress.

Ongoing development centers on integrating machine learning into serverless scheduling, aiming for systems that proactively respond to evolving security landscapes. This involves training algorithms to analyze real-time threat data – identifying patterns indicative of attacks – and dynamically adjusting function placement and resource allocation. Such adaptive scheduling goes beyond static configurations, enabling the system to prioritize security without sacrificing performance; functions at high risk of compromise can be isolated or replicated, while resource contention stemming from malicious activity is mitigated through intelligent distribution. Ultimately, this research strives to create serverless architectures capable of autonomous self-defense, continually learning and optimizing to maintain resilience against increasingly sophisticated threats.

The envisioned future of serverless computing centers on the development of a truly self-healing ecosystem. This isn’t merely about rapid recovery from failures, but proactive adaptation to maintain both security and performance under dynamic conditions. Such a system would leverage continuous monitoring and analysis of runtime behavior, identifying anomalies and potential threats before they escalate. Automated responses, ranging from function redeployment to dynamic scaling and security policy adjustments, would then be triggered without human intervention. This adaptive capability relies on intelligently shifting workloads, isolating compromised functions, and preemptively reinforcing defenses, ultimately creating a resilient infrastructure capable of autonomously preserving service levels even amidst evolving threats and fluctuating demands.

The pursuit of comprehensive serverless security often leads to architectures of unnecessary intricacy. Kumo, as described in this work, attempts to disentangle the genuine threats from the imagined ones-a crucial step towards pragmatic defenses. It’s a reminder that understanding the fundamental vulnerabilities, particularly those stemming from co-location and scheduling, is paramount. As Donald Davies observed, “Simplicity is a prerequisite for reliability.” The simulator’s focus on scheduler-driven isolation risks, distinct from generalized resource exhaustion, embodies this principle. They called it a framework to hide the panic, but a clear understanding of these basic risks is a far more effective strategy than layering complexity upon complexity.

Where Do We Go From Here?

The construction of Kumo, while a necessary exercise in clarifying the contours of serverless security, reveals more about what remains unknown than what has been solved. The simulator isolates scheduler-driven risks from simple resource exhaustion, a distinction frequently blurred in prior work. Yet, this very act of isolation exposes the inadequacy of addressing these problems in isolation. A scheduler that perfectly mitigates co-location attacks is still vulnerable to a sufficiently motivated attacker exploiting platform-level weaknesses-a truth often obscured by the complexity of real-world systems.

Future work must resist the temptation to build ever-more-detailed simulations. Such efforts, while impressive, merely shift the problem-obscuring fundamental limitations under layers of abstraction. Instead, attention should be directed toward formally defining the boundaries of isolation achievable in serverless environments. What guarantees, if any, can be provided beyond probabilistic assurances? Can resource contention, seemingly an unavoidable consequence of shared infrastructure, be framed not as a bug, but as a feature-a means of enforcing fairness or incentivizing efficient code?

The pursuit of perfect security is a fool’s errand. A more fruitful path lies in acknowledging the inherent trade-offs between security, performance, and cost, and in developing tools that allow platform operators to make informed decisions-not based on the illusion of absolute safety, but on a clear understanding of the risks involved. If the problem cannot be stated simply, then the solution will inevitably be flawed.

Original article: https://arxiv.org/pdf/2603.19787.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-03-24 00:13