CVE Tools

Google finds AI agent writing and launching exploits with no human in the loop

Help Net SecurityBy Sinisa Markovic

Our summary

Google's Threat Intelligence Group and Mandiant report that attackers are increasingly handing entire phases of intrusions to AI agents, including one case where exploit code was generated and deployed without any human involvement. Researchers traced this to an exposed staging server running a multi-agent system called exploit_agent, which compared application code before and after patches to spot flaws, write exploits, and fire them off automatically; the server also contained brute-force and webshell scripts with Russian-language comments, and telemetry pointed to interest in French media and education sites. The FBI separately warned that the open-source pentesting tool ARTEX has been weaponized against financial, government, defense, healthcare, and energy organizations since August 2026, alongside similar misuse of CAIRN, CyberStrikeAI, PentAGI, and HexStrike AI by both financially motivated and suspected China-linked actors. Google also documented LLMJacking incidents where stolen cloud credentials racked up over $1 million in AI compute abuse, a North Korean cluster (MIDNIGHT NEPTUNE) planting a SOMBERMEME backdoor via tampered Claude Code settings, and influence operations using AI to generate humanlike fake personas and manage over 100,000 automated social media accounts.

Read at Help Net Security

Help Net Security publishes this story on its own site; we link to it rather than reprint it.

Worried this affects your company?

Discuss a security assessment of your internet-facing systems. Scope agreed before testing.

Check my exposure