How Nation-State Hackers Are Using AI Agents to Run Fully Autonomous Cyberattacks

Robert McArdle, TrendAI's Director of Cybercrime Research, described AI as having become "a teammate embedded in the operation itself" rather than a side tool . The report covers activity observed from January to June 2026, targeting organizations in Australia, New Zealand, the United States, Ukraine, and other regions .

Autonomous AI Reconnaissance and Lateral Movement

In one documented case, a China-aligned group deployed an AI agent that independently conducted reconnaissance and lateral movement inside a target network without direct human instruction during those phases . McArdle warned that defenders must now assume "the adversary on the other end of an intrusion may not be a person typing commands, but a system executing a plan" .

This marks a critical shift: AI agents are no longer just assisting human operators but are acting as autonomous participants in attack chains .

'Vibe Coding' for Malware Refinement

China-aligned actors used generative AI to iteratively build and refine malware through conversational prompting — what TrendAI calls "vibe coding." Attackers treated AI like a developer pairing partner, generating and improving exploit code through natural language interactions, dramatically accelerating the malware development cycle .

Rapid Zero-Day Exploitation

The report documents how multiple groups weaponized vulnerabilities far faster than standard patching cycles can address:

• Russia-aligned Pawn Storm opened 2026 with an Office zero-day vulnerability, targeting Ukrainian government, defense, and wartime-aid organizations .
• Iran-aligned Earth Vetala scanned for a newly disclosed Ivanti vulnerability within days of its public disclosure .

These timelines illustrate what TrendAI calls "shrinking defense windows" — known and zero-day vulnerabilities are being weaponized within days of disclosure, compressing the time defenders have to patch. Organizations relying on monthly or quarterly patching cycles are increasingly exposed before fixes can be applied .

Physical-World OT Tampering

Iran-aligned actors conducted hands-on attacks against internet-exposed operational technology (OT) in the United States, tampering with fuel-tank gauge systems — equipment with real-world safety implications . This demonstrates that AI-augmented APT groups are not limited to digital targets but can directly affect physical infrastructure.

North Korea Supply-Chain Poisoning

DPRK-aligned groups incorporated commercial AI tools into their operations and poisoned a widely used software package to compromise downstream developers through the software supply chain . This approach allows attackers to infect many targets through a single compromised dependency.

ADINT: Malware-Free Surveillance via Ad Auctions

Beyond AI-specific tactics, the report also highlights "ADINT" (advertising intelligence), a malware-free tracking method that harvests location and device data from real-time online ad auctions for surveillance purposes . Because this method leaves no malware on the target device, it is exceptionally difficult to detect.

The Bottom Line for Defenders

The H1 2026 APT Activity Roundup makes clear that the threat landscape has fundamentally changed. AI is no longer an accessory to cyberattacks — it is the engine. Security teams must assume that adversaries can operate at machine speed, autonomously execute attack phases, and weaponize vulnerabilities faster than traditional patching can respond. The era of the human-only attacker is over.