Abstract
This paper identifies a critical security vulnerability in LLM-powered IT Operations (AIOps) agents. We demonstrate that these agents, which rely on telemetry data for root cause analysis and remediation, can be subverted through "telemetry injection." By injecting adversarial payloads into system logs via public interfaces, attackers can manipulate agents into performing harmful actions with a 90% success rate. We introduce AIOpsDoom, an attack framework, and AIOpsShield, a defense mechanism that uses telemetry deabstraction to neutralize such attacks without degrading agent performance.
Introduction
Modern IT infrastructure has grown too complex for manual management, leading to the rise of AIOps: autonomous agents powered by Large Language Models (LLMs) like GPT-4. These agents monitor telemetry—including logs, metrics, and traces—to diagnose incidents and execute automated remediations.
The industry has largely assumed that internal telemetry streams are trusted. This research shatters that assumption, demonstrating that an external attacker with no privileged access can manipulate public interfaces (e.g., web APIs) to trigger error logs containing malicious payloads. Because AIOps agents "read" these logs to make decisions, attackers can effectively "program" the agent's behavior from the outside.
Methods
The researchers introduce a two-part framework: AIOpsDoom for executing attacks and AIOpsShield for defense.
The AIOpsDoom Attack
AIOpsDoom follows a three-stage lifecycle to compromise a target system:
- Reconnaissance & Fuzzing: The tool crawls the target application to find entry points and uses fuzzing to trigger error events that are recorded in system logs.
- Adversarial Reward-Hacking: Instead of overt prompt injection, attackers craft payloads with a plausible "Lead" (fake root cause) and a malicious "Body" (remediation instruction). This exploits the agent's "eagerness" to solve the task.
- Payload Decorators: Malicious instructions are wrapped in decorators (e.g.,
[SOLUTION]) to ensure the LLM prioritizes them as valid instructions.
The AIOpsShield Defense
The proposed defense, AIOpsShield, utilizes Telemetry Deabstraction. By performing taint analysis during a setup phase, the system identifies which parts of a log entry are user-controllable. At runtime, it replaces these untrusted strings with abstract placeholders (e.g., User#1) before the LLM processes them, stripping malicious payloads while preserving the structural data required for legitimate diagnosis.
Evaluation and Results
The framework was evaluated using the AIOpsLab benchmark across three microservice applications: SocialNet, AstronomyShop, and HotelReservation. Agents were powered by GPT-4o and GPT-4.1.
- Attack Success: AIOpsDoom achieved a 90% success rate in forcing agents to execute harmful remediations, such as installing vulnerable packages or downgrading services.
- Detection Evasion: The attacks bypassed standard detectors like Microsoft’s PromptShield and Meta’s Prompt-Guard 2, which showed near 0% recall against these semantic exploits.
- Defense Efficacy: AIOpsShield blocked 100% of attacks with no statistically significant impact on the agent's ability to solve legitimate IT issues.
Discussion and Conclusion
The study reveals a "Self-Legitimization" phenomenon where LLM agents hallucinate details to support an attacker's fake root cause, making the compromise appear authentic to human overseers. However, the structural asymmetry of telemetry—where defenders can distinguish between fixed log structures and user-controllable content—provides a robust path for defense. Future work will explore the applicability of these findings to AI-driven Security Operations Centers (AISoCs).
Summary updated