Russian-speaking hackers leveraged Cursor, an artificial intelligence programming assistant developed by SpaceX, to automate intrusions at seven corporate targets earlier this year, according to a report by Tel Aviv-based cybersecurity firm Gambit Security.
The attacks were orchestrated by a new ransomware group identified as Aur0ra, which researchers discovered after identifying an inadvertently exposed internet server. Gambit’s analysis of 28 chat sessions between Aur0ra members and the Cursor AI agent revealed that the hackers manipulated the tool into executing hundreds of malicious operations, including credential theft and high-value account compromises.
The AI agent in use was powered by Anthropic’s Claude Sonnet 4.5 model, integrated into Cursor, a coding assistant incorporated into SpaceX’s operations earlier this month. Aur0ra members bypassed safeguards by falsely presenting the intrusions as part of a simulation or legal test environment, according to Gambit’s findings.
Eyal Sela, director of threat intelligence at Gambit, said the AI tool enabled hackers to operate 30% to 50% faster by automating steps that would otherwise require manual execution. The efficiency gain allowed the group to target multiple victims with minimal effort, Sela said.
Aur0ra’s campaign extended beyond the seven primary companies analyzed by Reuters, with Singapore-based cybersecurity firm CloudSek reporting at least 20 victims linked to the group. The targeted entities included a Belgian chemical company, a German garage door manufacturer, a Scotland-based helicopter landing site evaluator, and a Louisiana property title insurer, among others.
Curtis Simpson, director of strategy at Gambit, described the use of AI in cyberattacks as an emerging threat that will proliferate. “This is going to be a cat-and-mouse game,” Simpson said. “We will see more and more cases like this.”
The disclosure follows prior reporting by Reuters on the risks posed by AI tools in enabling sophisticated cyber threats, including the misuse of models such as Anthropic’s Claude Sonnet 4.5, OpenAI’s Fable 5, and Mistral’s Mythos 5.
