Key Takeaways

  • Google found AI agents combining vulnerability research, exploit creation and execution into connected workflows.
  • Systems including ARTEX, CAIRN, CyberStrikeAI and PentAGI could give defenders less time to patch exposed assets.
  • Faster remediation, phishing-resistant authentication and carefully bounded defensive automation can reduce business risk.

Google has documented how malicious actors are progressing from AI-assisted tasks to agentic workflows. Rather than relying on AI for an isolated coding task or a convincing phishing email, attackers are beginning to connect specialized agents across several stages of an intrusion.

Human operators still select targets, establish objectives and decide how to monetize access. Execution, however, can become far more automated. Agents can examine software updates, identify potential weaknesses, create exploit code, test results and adjust their approach with minimal human involvement.

Research from the Google Threat Intelligence Group and Mandiant illustrates how quickly that model can move. The Google Threat Intelligence Group (GTIG) observed a compromised cloud resource being used to plan, build and execute a mass credential-harvesting campaign in under six hours. That establishes a demanding benchmark for businesses whose incident workflows still depend on a ticket moving between several manual teams.

One system identified by Google, called exploit_agent, monitored software patches for evidence of vulnerabilities, inferred how those weaknesses could be exploited, generated potential exploits and attempted to execute them. The multi-agent system performed those activities without human intervention.

Patch analysis itself is not new. Attackers have long compared software versions to determine what a vendor fixed, then targeted organizations that had not yet installed the update. What changes with exploit_agent is the repeatability and potential scale. The process can keep running across numerous products and exposed applications.

The surrounding vulnerability environment is already accelerating. Google reported that vulnerability disclosures and exploitation nearly doubled from January through August 2026. Average monthly zero-day exploitation increased from eight cases in 2025 to 11 in 2026, then reached 22 in August.

Zero-days attract attention, but ordinary security gaps create broader commercial risk. Frameworks such as ARTEX, CAIRN, CyberStrikeAI, PentAGI and HexStrike AI can support activities ranging from reconnaissance and vulnerability testing to attempted exploitation. An unpatched server, forgotten development environment or overprivileged service account does not need to involve a novel technique. It only needs to remain discoverable.

The issue extends beyond Google’s findings. ENISA analyzed 4,875 incidents occurring from July 2024 through June 2025 and reported increasing AI use by malicious groups. ENISA also highlighted the other side of the equation: enterprise AI adoption is creating additional systems, dependencies and interfaces that defenders need to monitor.

Speed changes the economics of attack. Automation lets an operator probe more targets, repeat tests and abandon unproductive paths without spending the same amount of analyst time. It also shrinks the period between a patch revealing a vulnerability and exploitation beginning in the wild. A seven-day emergency patch process becomes inadequate if an offensive agent starts testing the update within hours.

Email remains part of that picture. Hornetsecurity found that malware-laden email volume rose 131% year over year in 2025, while 77% of CISOs considered AI-generated phishing a serious emerging threat. Agentic capabilities could connect that initial lure to credential validation, cloud discovery and follow-on activity, making campaigns more persistent even when individual attempts fail.

Businesses can respond by tightening operational basics. An accurate inventory of internet-facing systems helps teams identify what is reachable before an automated scanner does. Vulnerability prioritization should consider active exploitation, exposure and business impact, rather than relying solely on severity scores. Unused services and abandoned cloud resources warrant particular attention.

Identity controls provide another barrier. Phishing-resistant multifactor authentication, limited service-account permissions, credential rotation and independent approval for payment or account-reset requests constrain what an automated workflow accomplishes after gaining access. Synthetic audio or video should not serve as the sole evidence of identity.

Defensive AI can help triage alerts, summarize investigations and recommend containment steps. Broad autonomous permissions introduce their own exposure, though. Organizations can bound security agents by limiting accessible systems, logging consequential actions and requiring human review for disruptive decisions. Google’s warning is ultimately about operating tempo: attackers are assembling reusable workflows, and defenders will need faster remediation without surrendering oversight or accountability.