Skip to main content

AI tool turns software patches into exploits without human involvement

Google says an exposed server held an autonomous exploit tool and stolen data suggesting attacks on French websites.

By Mei Ling Tan

Computer monitor displaying software code in a cybersecurity workspace
Photo: TNW (Financial Times)

Google said it has observed attackers using AI to identify software vulnerabilities and exploit them without human involvement for the first time. Its latest threat report, produced by the Google Threat Intelligence Group (GTIG) and Mandiant, describes a tool that independently develops and tests exploits by examining software patches.

Investigators discovered the tool, named exploit_agent, on an exposed staging server. According to GTIG, it supports several frontier AI models as well as locally operated models, comparing versions of widely used web software to identify potential weaknesses revealed by updates.

The same server contained stolen victim information and scripts for password guessing, deploying webshells and deleting those installed by competing attackers. GTIG said the victim data pointed to targeting of French websites in media, entertainment and education. Comments in a configuration file were written in Russian.

Google said patched vulnerabilities can present easier opportunities for AI because the material to examine is limited, including code changes, release notes and published research. The report cited two PaperCut print-server flaws that attackers combined in late August. Researchers and security company Huntress developed working exploits using AI within a day of disclosure.

GTIG also reported that financially motivated attackers and suspected Chinese espionage groups are using open-source AI hacking frameworks in active operations. Targets of the CAIRN framework included the US aerospace sector. Google said these tools currently generate conspicuous activity that helps defenders detect them.

Separately, the report described stolen cloud and AI credentials being used to charge model operations to victims’ accounts. Google said it had disabled accounts and assets associated with misuse of its Gemini models.

Keep reading: Tech News

All Tech News stories

Latest in Technology

All latest
  1. Tech News

    Noeri announces US$110 million funding round

  2. Tech News

    Apollo to embed six evaluators at OpenAI for safety tests

  3. Tech News

    FAA router investigation follows outage that halted U.S. flights

  4. Tech News

    SK Hynix plans supplier support package with shared chip testing site

  5. Gadgets

    Microsoft positions Project Helix as a console with PC game support

  6. Tech News

    Nvidia’s interim CEO expected to scale back AI financing