Evals

This module is a test orchestrator, not an obvious malware container. However, it executes externally influenced commands: it interpolates workingDir, a possibly user/config-provided command (params.test_command), and testFile strings directly into a Bun $ execution sink without visible validation/quoting/allowlisting. If an attacker can influence grading parameters or context, this creates a substantial command-injection / arbitrary command execution risk. Additionally, it returns truncated stdout/stderr from executed commands, which can leak sensitive environment/test output to upstream consumers.

This module does not itself implement malware (no network exfiltration, persistence, credential harvesting, or obfuscated payloads are present in the snippet). However, it is a high-impact execution harness: it directly executes arbitrary shell commands sourced from configuration via Bun’s shell template and returns captured stdout/stderr/exception strings to callers. If upstream config or workingDir inputs are not strictly trusted/validated, this creates a significant command execution and potential information disclosure risk. Additionally, grading is based on heuristic text matching rather than exit codes, which can be manipulated by command output formatting.