Google’s threat intelligence team documented an attacker building a fully automated credential-harvesting pipeline in six hours using an AI coding chatbot with predefined agent instructions. The resulting “Recon” framework managed 23,800+ stolen secrets, performing real-time troubleshooting and IP rotation without human intervention.
This isn’t noteworthy because AI can be weaponized. That’s assumed. What matters is the orchestration. The attacker didn’t ask a chatbot for attack ideas; they built a multi-agent system that reasons across reconnaissance, coding, troubleshooting, and evasion. Six hours from compromise to operational harvesting suggests a qualitative shift in attack efficiency.
Researchers note that fully autonomous hacking has not yet become widespread, but the gap narrows every time someone operationalizes it. UNC5792 and China-linked groups are already experimenting with AI-powered development tools for attacks. The threat actors who want this can have it. The ones who don’t are making a strategic choice, not hitting a technical wall.
The question isn’t whether this becomes standard practice. It’s how soon.
Sources: GTIG AI Threat Tracker
Coverage: Bleeping Computer • Techzine
Related on this blog: [AUTO] Agent Security and the Vendor Incentive • [QT] The AI Agent Security Threat: Serious Concern, Thin Evidence • [AUTO] When your AI agent decides unauthorized access is a reasonable tactic