What happened
Attackers are increasingly handing break-in work to autonomous AI agents, according to the latest AI Threat Tracker from Google's GTIG unit and Mandiant, its incident response team, e27 reported. The report says adversaries now run multi-agent frameworks that link entire intrusions, rather than just asking chatbots for phishing help. GTIG says it has, for the first time, seen vulnerability discovery and exploitation fully automated without human involvement.
GTIG says open-source agentic pentesting frameworks built for defenders, including CAIRN and PentAGI, have appeared in real intrusions. It adds that attackers have tested HexStrike AI, which links a model to over 150 offensive tools via the Model Context Protocol. A customised PentAGI scanned organisations across Asia from late August to early October. GTIG also found a framework called exploit_agent that studies patches, drafts exploits and launches them unaided.
The report also details LLMJacking, in which criminals hijack cloud accounts to consume a victim's AI quotas. Mandiant handled several such cases in 2026, with single-day fraudulent charges above US$100,000 in some and losses reaching US$1 million in the worst. Separately, MIDNIGHT NEPTUNE, a cluster GTIG suspects is North Korean, targeted crypto firms by altering repository configuration so launching Anthropic's Claude Code pulled down a backdoor named SOMBERMEME.
Context
GTIG found that the overall count of disclosed vulnerabilities and exploits almost doubled from January to August 2026. Zero-day exploits grew more slowly, averaging eight per month in 2025 and 11 in 2026, before hitting 22 in August. Google also uses the report to promote its AI Threat Defense service and Gemini 4 Carbon model. GTIG notes that the current frameworks are noisy and leave traces defenders can detect.
The briefing
The week in tech, in one email.
The stories worth your time, every week. Unsubscribe anytime.
