Agentic AI: Chinese APT Groups Are Automating Cyberattacks

AI Agentica: i Gruppi APT Cinesi Automatizzano gli Attacchi

The Google Threat Intelligence Group (GTIG) has published a report marking a significant turning point in how threat actors are leveraging agentic AI. For the first time, Chinese-nexus cyber-espionage groups have moved well beyond simple experimentation. They are now building autonomous workflows capable of executing attacks in an automated, scalable fashion.


Agentic AI: What GTIG Uncovered

From Prompting to Operational Autonomy

Until recently, malicious actors were using large language models in fairly rudimentary ways — generating phishing text or probing for information about known vulnerabilities. That picture has changed dramatically.

GTIG has documented a Chinese-nexus espionage group using Gemini to develop an automated penetration testing framework. This system does far more than generate code. It observes the state of a target, reasons about the next steps, and executes them autonomously. In practice, this is an offensive agent.

One important distinction is worth making. GTIG describes capability development and experimentation — no specific breach with identified victims or quantified damage has been confirmed. Nevertheless, the direction of travel is clear and deeply concerning.

Credential Harvesting in Six Hours

In parallel, GTIG and Mandiant documented a multi-agent credential theft campaign. Operating from compromised infrastructure, attackers reportedly harvested thousands of credentials in roughly six hours.

This figure illustrates the potential speed of AI-assisted attacks. Scale is no longer constrained by the attacker’s human resources. With automation, the marginal cost of each additional attack approaches zero.


Threat Actor Profiles and At-Risk Sectors

Chinese Espionage and Commercial AI

The report attributes the primary activity to a Chinese-nexus threat actor. GTIG has separately linked AI use to groups including APT27 and APT41, though these attributions should not be automatically applied to the specific incident described in the report.

What does emerge clearly is a consistent pattern of behavior. Threat actors are leveraging commercial AI models — including Gemini, Claude, and Codex — to generate exploits, craft spear-phishing lures, develop post-exploitation tools, and automate reconnaissance.

The Most Exposed Sectors

No sector is immune in this environment. That said, technology, government, cloud providers, and organizations holding valuable intellectual property remain the recurring targets of Chinese espionage. The pattern is systemic, not confined to any single industry.

Security teams therefore need to think beyond more convincing phishing emails. They must prepare for attacks that are faster, more scalable, and far harder to detect with traditional controls.


Defending Against Offensive Agentic AI

Priorities: Identity, Cloud, and Rapid Detection

GTIG, alongside CISA and NIST guidance, points clearly to the defensive priorities. Identity is the new perimeter. Attackers are using AI to steal and abuse credentials, tokens, and sessions — meaning controls must be both stronger and more granular.

Key operational recommendations include:

  • Phishing-resistant MFA (FIDO2 keys) across all privileged accounts
  • Conditional access policies based on device, user, location, and risk score
  • Short-lived tokens and rigorous protection of service account credentials
  • Privileged Access Management with just-in-time administration
  • Centralized logging covering identity, cloud, APIs, endpoints, and data access

Automated Detection and Response

Organizations must also invest in detection capabilities specifically tailored to AI-assisted attacks. Key signals to monitor include anomalous OAuth grants, unusual token usage, impossible travel events, and mass cloud resource creation.

At the same time, incident response playbooks need to evolve. Agentic attacks iterate rapidly. Security teams must be able to automatically contain compromised accounts, tokens, endpoints, and cloud workloads — reserving human approval for high-impact decisions only.

Governance over internal generative AI use deserves equal attention. Organizations should establish clear policies on employee use of AI tools, including restrictions on sending sensitive data to external models and monitoring of AI-generated code or content.


Conclusions

Agentic AI is no longer a future risk. It is a documented reality, confirmed today by GTIG. Chinese espionage groups are testing and deploying autonomous offensive workflows. A threshold has been crossed.

For CISOs and security leaders, the message is straightforward. Hardening identity and cloud environments is no longer optional — it is the number-one priority for withstanding the next generation of automated attacks.


Sources:

Source: Original article


The emergence of agentic AI attacks from Chinese APT groups makes real-time threat intelligence sharing between organizations in the same sector more urgent than ever. IsacChain addresses this need by offering a secure, blockchain-based platform for sharing verified threat intelligence — enabling automated NIS2 compliance through immutable reports and audit trails. In an environment where attacks move at machine speed, the ability to correlate indicators of compromise across multiple organizations in a trusted, traceable way becomes a decisive defensive advantage. Discover how IsacChain can help your organization at www.isacchain.com