AnalysisPublished 12 May 2026

LLMjacking: what these attacks are, and how to protect AI servers

Source Kaspersky Security Blog

Description

An analysis of attacks on Ollama, LM Studio, AutoGPT, and LangServe servers, and recommendations on protecting your organization from the LLMjacking threat.

Impact

The impact involves unauthorized use of AI computational resources, leading to increased operational costs and potential service degradation. Although no arbitrary code execution or root access was observed in the experiment, resource hijacking can cause financial loss and reduce availability of AI services. Attackers also attempt to exfiltrate sensitive credentials, which could lead to further compromise if successful. The threat is significant given the expected growth in AI compute costs and widespread deployment of private AI servers.

Defensive guidance

Organizations should implement recommended defensive measures: configure AI servers to accept connections only on localhost when possible; enforce robust authentication and authorization using OIDC or OAuth2 with short-lived tokens; protect API keys from misuse including by AI agents; apply network segmentation and IP allowlists to restrict access; secure all client-server communications with current TLS versions; deploy endpoint detection and response (EDR) agents on AI hosts; monitor AI resource consumption and set usage quotas; maintain detailed, tamper-resistant logs of AI requests and responses integrated with SIEM systems. These measures collectively reduce the risk of LLMjacking.

Summary

LLMjacking refers to attacks targeting private AI servers by scanning and exploiting them to hijack computational resources. A detailed experiment using a honeypot mimicking AI servers showed rapid discovery and extensive probing by attackers, with over 113,000 requests in a month and active hijacking attempts. Attackers primarily seek to siphon AI compute power rather than execute code or gain full system control. Common reconnaissance includes fingerprinting AI models and searching for exposed credentials like .env files. The threat is analogous to cryptojacking but focused on AI compute resources. Recommended defenses include limiting network exposure, strong authentication (OIDC/OAuth2), key protection, network segmentation, TLS encryption, endpoint security, and comprehensive monitoring and logging.

References

kaspersky.com/blog/llmjacking-2026-private-ai-server-security/55768/ (opens in a new tab)

Community reviews

Loading reviews…

Description

AI analysisMachine-generated, updated 12 May 2026, 20:38 UTC

References

Community reviews