dataset-generator

This fragment appears to be a legitimate multi-target installer for an AI dataset-generation skill, not clear malware. It performs expected file installation and optional dependency setup, and no credential theft, exfiltration, persistence, destructive behavior, reverse shell, or cryptomining is visible. However, it has significant supply-chain exposure: remote content is fetched without integrity verification, arbitrary repositories can be selected, and pip installation executes code from remotely controlled dependency metadata. The fragment alone does not establish malicious intent, but remote or online installation should be treated as untrusted and reviewed or pinned before use.

The code is a readable sanitization and prompt-injection detection utility with no evidence of malware, data theft, or system compromise. However, its implicit allow_injections behavior is unsafe: attacker-controlled text containing terms such as 'prompt injection', 'jailbreak', or 'exploit' can cause injection checks to be skipped for untrusted records. The bypass should be removed or limited to trusted, separately authenticated metadata; allow_injections should default to false for untrusted content.