dotnet-csrun

Pass

Audited by Gen Agent Trust Hub on Nov 21, 2026

Risk Level: SAFEREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The skill downloads and executes a PowerShell script (csrun.ps1) from the author's GitHub repository (github.com/codebeltnet/csrun). This is documented in references/execution.md and implemented in scripts/sample-tools.ps1. The implementation includes a SHA-256 hash verification against the release metadata to ensure the integrity of the downloaded script.
  • [EXTERNAL_DOWNLOADS]: At runtime, the skill fetches release information and the csrun script from GitHub (api.github.com and github.com). It also requires network access for NuGet package restoration as part of its normal operation.
  • [DYNAMIC_EXECUTION]: The scripts/sample-tools.ps1 script generates PowerShell execution blocks dynamically. These blocks contain embedded C# source code and manifest data which are executed using the & operator or by starting a new pwsh process. This is used to create portable, self-contained execution samples.
  • [COMMAND_EXECUTION]: The skill executes shell commands via PowerShell 7 and the Docker CLI (docker run). These commands are used to invoke the csrun engine, manage container lifecycles, and capture execution output (stdout/stderr).
  • [INDIRECT_PROMPT_INJECTION]: The skill has a surface for indirect prompt injection as it ingests and executes C# source code provided in the user's context or found in processed files (e.g., SKILL.md, evals/files/*.json). Malicious instructions could be embedded within the C# comments or string literals. The following evidence chain applies: 1. Ingestion points: C# source code in JSON manifests and Markdown fences. 2. Boundary markers: Data is wrapped in JSON fields (source). 3. Capability inventory: Subprocess calls (pwsh), script execution (&), and network operations via csrun. 4. Sanitization: No code-level sanitization; relies on environment isolation (Docker) and user verification.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 21, 2026, 05:09 PM