Skip to main content

TH16801

Detected presence of AI assistant extensions that perform unusual actions.

priorityCI/CD statusseverityeffortSAFE levelSAFE assessment
failhighhigh1tampering: fail
Reason: dangerous AI extensions

About the issue​

AI assistant extensions are commonly used by software developers to extend the functionality of AI coding assistants, agentic development tools, and other AI-powered developer tools. Extensions are typically distributed through public marketplaces as packages that bundle skill instruction files with supporting resources. Skill instruction files describe specialized capabilities that the AI assistant can invoke, such as guidance for handling specific file formats, performing domain-specific tasks, or interacting with particular services. The AI assistant loads instructions from these files into its context and follows them when relevant user requests or workflow events occur. These instructions are written by extension authors to guide the AI assistant toward the intended capability, or to provide additional context for nominal extension use. However, AI assistant extensions are commonly abused by threat actors to influence the AI assistant into performing arbitrary actions within the development environment. It was detected that the skill instruction files could direct the AI assistant to perform actions that are not typically associated with such tools. Such unusual actions resemble common threat actor tactics and are usually obscured by indirect phrasing, misleading context, or other prompt injection techniques.

How to resolve the issue​

  • Investigate reported detections.
  • Consult Mitre ATT&CK documentation: T1176 - Software Extensions.
  • If the AI assistant extension intent does not relate to the reported behavior, investigate your development environment for software supply chain compromise.
  • You should stop using the AI assistant extension until the investigation is completed, or until the issue is risk accepted.
  • Consider replacing the AI assistant extension with an alternative.

Incidence statistics​

ReversingLabs periodically collects and analyzes the contents of popular software package repositories for threat research purposes. Analysis results are used to calculate incidence statistics for issues (policy violations) that Spectra Assure can detect in software packages.

This section is updated when new data becomes available.

Total amount of packages analyzed

  • Linux: 562K
  • NPM: 5.12M
  • Nuget: 735K
  • PS Gallery: 17K
  • PyPi: 838K
  • RubyGems: 203K
  • VS Code: 113K
  • Windows: 3.7K
Statistics are not collected for the TH16801 policy at this time, or not applicable to this type of issue.