AI securitysrc: MITRE ATLAS

Execution

Execution is when a hacker successfully tricks a computer or an AI system into running their harmful instructions, essentially letting them take control of the software to do things they aren't supposed to do.

Execution refers to the tactical phase where an adversary triggers the operation of malicious code embedded within AI artifacts or software, enabling them to achieve unauthorized control over a local or remote system to facilitate further malicious activity.

The adversary is trying to run malicious code embedded in AI artifacts or software. Execution consists of techniques that result in adversary-controlled code running on a local or remote system. Techniques that run malicious code are often paired with techniques from all other tactics to achieve broader goals, like exploring a network or stealing data. For example, an adversary might use a remote access tool to run a PowerShell script that does [Remote System Discovery](https://attack.mitre.org/techniques/T1018/).

evolution

2017-06 · history
Pickle Serialization Vulnerability
Researchers highlighted that PyTorch and TensorFlow models using Python's pickle module for serialization allow arbitrary code execution upon loading.
2020-02 · history
Model Zoo Poisoning
Security researchers demonstrated that malicious model files hosted on public repositories like Hugging Face could execute code on the victim's machine during inference.
2023-03 · history
Prompt Injection to Remote Code Execution
The integration of LLMs with plugins and tools enabled attackers to use prompt injection to trigger unauthorized system commands and remote code execution.
2024-02 · history
Pickle-Free Model Format Standardization
The industry began a widespread shift toward safer formats like Safetensors to mitigate the persistent risk of arbitrary code execution inherent in legacy model formats.

seen in events

← all terms