OpenAI caught Russian hackers using ChatGPT to tweak malware line-by-line

Russian-speaking hackers used a clever but ultimately self-defeating trick to build malware with ChatGPT: they burned through disposable accounts, asking for one small code fix per conversation before ditching the account entirely. OpenAI’s threat intelligence team, which detailed the operation in a June 2025 report, nicknamed the campaign “ScopeCreep” after the hackers’ weapon of choice—a trojanized version of a legitimate gaming crosshair tool called Crosshair-X.

The operational security play was simple. Sign up with a temporary email, paste in a snippet of broken Go code, get a fix, and abandon the account. Rinse and repeat. The hackers used this method to stitch together a multi-stage Windows implant capable of escalating privileges, stealing browser credentials and session tokens, and phoning home to a command-and-control server over HTTPS on port 80. They even asked the model to write PowerShell commands in Go that would programmatically add Windows Defender exclusions—a classic AV evasion move.

Here’s the irony: the very behavior that the hackers thought would hide them—fragmenting their workflow across dozens of accounts—created a pattern that OpenAI’s automated abuse detection was built to catch. The company spotted the cluster, banned every associated account, and coordinated with the code-hosting provider to yank the malicious repo. One representative exchange shows the threat actor struggling with an SSL/TLS error and asking ChatGPT to debug it, mapping neatly to the “LLM Aided Development” technique in the LLM ATT&CK framework.

The payload itself wasn’t groundbreaking. It used Themida for packing, DLL side-loading for stealthy execution, and Telegram for victim notifications—techniques that have been in the wild for years. Some samples did appear on VirusTotal, but OpenAI says it saw no evidence of widespread distribution. What makes ScopeCreep notable isn’t the malware’s sophistication, but the operational model: an adversary treating a frontier AI model like a stack overflow that never sleeps, while inadvertently handing defenders a neon sign pointing straight at their infrastructure. It’s a dynamic we’re going to see a lot more of—attackers accelerating development with AI while simultaneously making themselves easier to fingerprint and disrupt.