Eigen RadarAI
Analysis

Google uncovers AI agents executing stages of live cyberattacks

Google’s security researchers have disclosed intrusions in which attackers used AI agents to scan targets and exploit weaknesses. One observed system compared software before and after patches, then prepared and ran attacks. The findings also cover stolen access to cloud AI services. Google describes faster operations accompanied by detectable traffic and code traces, rather than an inventory of every attack.

Artificial Intelligence··Evening
A security researcher uses a stylus to examine side-by-side software patch comparison bars, network traffic waveforms and agent node traces on a forensic display in a threat-intelligence laboratory.

Agents compare patched code and prepare attacks

Google Threat Intelligence Group, the company’s security research team, and incident-response specialist Mandiant have disclosed observations of AI agents in real intrusions. Attackers used frameworks that can coordinate several agents to scan targets and exploit weaknesses. The newly disclosed findings concern operations observed over preceding months, including a modified PentAGI platform used against organizations in Asia from late August to early October.[1], [2]

The researchers observed systems comparing application code before and after security patches to prepare exploitation of known vulnerabilities. CAIRN, another agent framework, was observed against targets including the US aerospace sector. The assessment also describes experimentation with CyberStrikeAI and HexStrike AI.[1]

Automated operations leave detectable traces

Google says autonomous execution can increase speed while generating recognizable traffic and code traces for investigators. Defenders also face model-generated exploits advertised as working that do not function. These nonfunctional examples add verification work during the response to newly disclosed vulnerabilities. Google’s findings distinguish the use of agent tools in an intrusion from an assumption that every part of every operation is autonomous.[1]

Stolen cloud access generates large unauthorized bills

Mandiant says some intrusions involving unauthorized use of cloud computing and language-model accounts had fraudulent charges exceeding one hundred thousand dollars in a day. The most severe accumulated abuse reached $1 million. Intermediaries sell access distributed across stolen credentials and computing resources from multiple victims. Google says it has disabled abusive accounts and shares findings with its DeepMind AI research team to improve safeguards against misuse.[1]

References

  1. News sourceGoogle Threat Intelligence GroupGoogle reports multi-agent tools being used in real intrusions↩1↩2↩3↩4
  2. News sourceHelp Net SecurityGoogle finds AI agent writing and launching exploits with no human in the loop↩