Google: AI turned patches into exploits
Google Threat Intelligence Group · first with no human in the loop
Security
A synthetic internet for prompt-injection tests
Written with AI assistance, reviewed by the ACN desk.
“Chrome is introducing the ‘mirror dimension’; a synthetic version of the web to test agents safely in realistic online environments, without creating real-world security risks.”
Thumbnail: Image: Google blog
Google’s Chrome team has built a synthetic copy of the web to test its AI browsing agents. Google’s security blog calls it a “mirror dimension” for agents.
The sandbox sends an agent’s traffic to a local proxy that serves generated websites. A multi-agent system builds those simulated sites. Google says the pipeline creates more than 1,000 test candidates every day per security engineer. According to the post, that is over 30x what one engineer could do by hand.
Google says pointing an agent at the public web during testing can be ineffective and irresponsible. It says agents can cause harm when given unfettered access. The team also uses the environment for red-teaming, or simulated attacks. Those tests target indirect prompt injection, where attackers hide instructions inside untrusted web content.
The post is by Aaron Selya and Charles Qian, software engineers on Chrome AI Security. Google calls the system a critical tool to keep new agent capabilities safe, reliable and resistant to threats.
“Chrome is introducing the ‘mirror dimension’; a synthetic version of the web to test agents safely in realistic online environments, without creating real-world security risks.”
Sponsor
Spot an error? Email corrections@aicodingnews.si. Fixes are logged on the corrections page.
Sponsor
Sponsor
Google Threat Intelligence Group · first with no human in the loop
Attackers used the open-source tool Cairn
Reuters cites internal posts; Meta has not responded
Poisoned agent-framework versions stole credentials
Sponsor
Sponsor