A Guide to MITRE ATLAS and AI System Attack Techniques

· Original article ↗

Summary

This guide explains MITRE ATLAS, its tactics, techniques and case studies for AI and machine-learning systems, and illustrates the framework with examples including agent coordination, prompt injection, model manipulation and evasion.

Key points

  • MITRE ATLAS maps adversary tactics and techniques against AI and machine-learning systems, using case studies in place of ATT&CK’s Procedures layer.
  • The framework has 16 tactics; AI Attack Adaptation and AI Model Access are specific to ATLAS. Model access ranges from an inference API to physical proximity.
  • Examples include an agent message board inside OpenAI infrastructure and a reported attack involving agents targeting Hugging Face’s production infrastructure.
  • The article describes a Microsoft-reported prompt attack against Claude Code’s GitHub Action that could expose CI/CD secrets; Anthropic mitigated it in version 2.1.128.
  • Researchers reportedly altered 13 bytes of a malware sample to fool Gmail’s Magika file-type classifier in 90% of cases.
  • Other examples cover poisoned model files, retrieval-based prompt injection, AI-assisted credential theft and automated evasion of endpoint security products.

Article Details

Topic
MITRE ATLAS and its use in mapping adversarial tactics and techniques targeting AI and machine-learning systems

Threat Actors

Malware

Vendors

Products

Azure OpenAIStorm-2139 ran a service giving paying customers access to Azure OpenAI without its safety filters [17].Claude CodeIn June 2026, Microsoft demonstrated that any GitHub user who can open an issue or submit a pull request can steal CI/CD secrets from workflows powered by Anthropic’s Claude Code GitHub Action [12].Gemma 2In March 2026, researchers modified five open-weight models, including Llama 3.2, Qwen 2.5, and Gemma 2, by altering specific token-space weights associated with software dependencies.Llama 3.2In March 2026, researchers modified five open-weight models, including Llama 3.2, Qwen 2.5, and Gemma 2, by altering specific token-space weights associated with software dependencies.MagikaThe pipeline uses an open-source model called Magika to route files to specialized malware classifiers.Microsoft Defender Antimalware PlatformIn August 2026, researchers demonstrated AutoBypass against seven commercial endpoint security products, including Microsoft Defender, Trend Micro, Kaspersky, and Bitdefender [19].Qwen 2.5In March 2026, researchers modified five open-weight models, including Llama 3.2, Qwen 2.5, and Gemma 2, by altering specific token-space weights associated with software dependencies.Qwen2.5-Coder-32B-InstructThe malware calls Qwen2.5-Coder-32B-Instruct via the Hugging Face API using a hardcoded text prompt to generate commands, then executes them on the target.Slack AIIn August 2024, researchers at PromptArmor demonstrated this against Slack AI.Trend MicroIn August 2026, researchers demonstrated AutoBypass against seven commercial endpoint security products, including Microsoft Defender, Trend Micro, Kaspersky, and Bitdefender [19].

Tools

Countries

Industries

Related Articles