NIK C2 malware can hijack victims' OpenAI keys for AI attacks

Cisco Talos has documented a new Windows command-and-control framework called NIK C2 that uses generative AI inside an active malware session. Talos says it observed the Go-based tool in the wild, making this more than a proof of concept.

The framework includes an OpenAI Chat Completions client. An operator can send arbitrary prompts through an infected computer and receive the model’s response from the same control channel. If the attacker does not supply an API key, NIK C2 can read the OPENAI_API_KEY environment variable from the victim’s machine. That lets the operator use the victim’s account for AI analysis while the victim absorbs the usage cost and their key appears in access logs. The malware defaults to GPT-4, but the operator can select another supported model.

NIK C2 also carries a broad set of conventional capabilities: credential dumping, keylogging, screenshots, camera and audio capture, Remote Desktop hijacking, proxying for lateral movement, security evasion and screen-locking ransomware. AI does not replace those functions; it compresses data theft and analysis into one workflow.

For developers and security teams, the immediate action is straightforward: avoid leaving API keys in broadly accessible environments, rotate any key exposed on a compromised host, and monitor model usage and spending for unexpected activity. An unusual OpenAI bill may now be an incident signal, not merely a configuration mistake.