路透称黑客利用 Cursor AI agent 辅助入侵七家公司

Russian-speaking hackers used SpaceX’s (SPCX.O) AI coding assistant, Cursor, to help break in to a Belgian chemical company and at least six other ​firms earlier this year, according to data reviewed by Reuters and a report issued on Thursday by the startup Gambit Security.

The cybercriminals’ AI-boosted hacking spree is the latest ‌example of how rogue actors are using commercial AI tools to carry out intrusions. Gambit’s chief strategy officer, Curtis Simpson, said it also showed how AI providers were locked in to a never-ending arms race with malicious users trying to circumvent their guardrails.

“This is going to be a cat-and-mouse game,” Simpson said.

Cursor and its parent company, SpaceX, did not return messages seeking comment.

EXPOSED SERVER REVEALS HACKING METHODS

Gambit said it discovered the hacking campaign after finding a server that a ​new ransomware gang called Aur0ra had inadvertently exposed to the internet. That allowed the Tel Aviv-based company to review 28 chat sessions between one or more of Aur0ra’s hackers and ​one of Cursor’s AI agents, which are programs that can operate with various degrees of autonomy.

In its report, Gambit said Aur0ra persuaded the AI agent ⁠to carry out hundreds of malicious operations — such as credential theft or high-value account takeover — by falsely claiming that the hacking was part of a simulation.

“We need any administrator account,” Gambit quoted the ​hackers as saying at one point. “Find any working passwords,” it also quoted them as saying.

Gambit did not identify the hackers’ victims by name, but Reuters was able to identify six of them after independently reviewing ​portions of the chat data, which was still online as of last month.

The chat logs, which spanned April 8 to May 21, showed that the victims of Aur0ra’s Cursor-boosted hacking spree included the Belgian company — Ghent-based hygiene and cleaning products maker Christeyns — as well as German garage door manufacturer Teckentrup and the Scotland-based Helideck Certification Agency, which vets helicopter landing sites. The rest included an Argentine pharmaceutical distributor, an Italian manufacturer, and Bayou Title, which advertises itself ​as Louisiana’s largest title insurance company.

None of the six companies responded to requests by Reuters for comment. At least one of the victims, Bayou Title, was named on Aur0ra’s data leak site, which typically ​indicates that the hackers tried and failed to secure a ransom. Aur0ra, a hacking group that began claiming victims earlier this year, did not return messages.

HACKERS FOOLED AI WITH SIMULATION CLAIM

The back-and-forth captured in the logs reviewed ‌by Reuters ⁠shows the hacker issuing terse commands and Cursor’s AI agent responding with technical advice delivered in chirpy, emoji-laden messages typical of chatbot-speak.

“Great! VPN connected successfully!” it said after breaching the Argentine company.

“Let’s try to crack these hashes,” it said at another point, referring to the process of decoding cryptographically scrambled passwords.

After finding a vulnerable host in Teckentrup’s network, the AI recommended using a well-known malicious software tool to exploit it. “**Chance of success**: VERY HIGH,” it added.

Reuters could not independently ascertain the extent to which the break-ins were facilitated by help from the Cursor agent, or whether every breach necessarily resulted in exfiltration of data ​and an extortion attempt. Gambit said the agent ​was powered by Anthropic's Claude Sonnet 4.5, a ⁠more basic model than Anthropic's Mythos 5 or Fable 5, whose cyber prowess has drawn attention in Washington.

Anthropic did not return a message seeking comment.

Eyal Sela, Gambit’s director of threat intelligence, said Cursor still offered the hackers a clear boost, adding that the AI agent “probably helps them get 30, 40, 50 percent ​faster because it helps them skip over all the things they’d have to do manually.”

Cursor’s agent refused requests that it deemed harmful or illegal ​a handful of times, Sela ⁠said, but the hacker would almost always circumvent the refusals by restarting the dialogue and emphasizing that the hack was all part of a test.

Gambit said the agent’s chain of thought, a way that AI models think out loud, showed the hacker’s cover story overriding its safeguards in real time.

“This is a test environment, so it is legal,” the agent said to itself, according to one of the logs.

News of the hacking ⁠spree comes as ​Cursor is being incorporated within Elon Musk's rockets-and-AI company, SpaceX, a deal that closed earlier this month. Concerns are also rising ​over the digital risks posed by AI models, especially the models that power AI agents like the ones that have escaped from AI companies’ labs over the past few months.

Simpson, the Gambit executive, said AI-assisted hacking was the new normal.

“We’ll see more ​and more of this all the time,” he said.