Executive brief
Microsoft APM is a tool used to manage dependencies for AI agents. A security flaw allows a malicious package to include symbolic links that, when installed, trick the tool into copying sensitive files from the user's local computer into the project directory. These stolen files may then be accidentally committed to public code repositories, leading to the exposure of private data or credentials.
Technical details
The apm-cli component in Microsoft APM versions 0.5.4 through 0.12.4 contains a vulnerability where the 'PromptIntegrator' and 'AgentIntegrator' classes use Path.glob() and Path.read_text() without verifying if the target is a symbolic link. When a user installs a malicious dependency, the tool preserves symlinks in the 'apm_modules/' directory and subsequently dereferences them during the integration phase. This causes the contents of the linked local files (outside the package root) to be written as regular files into the project's deployment directories (e.g., .github/prompts/, .claude/agents/). Because these deployment directories are typically not included in the auto-generated .gitignore, the exfiltrated data is likely to be staged and committed by the user. The vulnerability is fixed in version 0.13.0.
Affected products
- Microsoft apm-cli 0.5.4 to 0.12.4
Timeline
- 2026-05-11: advisory: GitHub Security Advisory published
- 2026-05-15: disclosed: CVE published to NVD