AI-powered cyber attacks slash response time to zero, experts warn

by Melania Watson

A Chinese state-sponsored group has used an artificial intelligence agent to automate most stages of a cyber attack, according to recent findings. The methods used compressed weeks of manual tradecraft into seconds, signalling a fundamental change in the speed and scope of offensive cyber operations.

Automation threat

The attack, referred to as the GTG-1002 campaign, saw the group exploit known vulnerabilities and orchestrate open-source tools using an AI agent based on Claude. For years, organisations have relied on a window of time between the discovery of a vulnerability and its exploitation. That window has now been reduced to zero, making traditional patching cycles much less effective.

During the campaign, the AI agent automated recon, exploit writing, lateral movement, and exfiltration. These actions, previously requiring human attackers days or weeks to accomplish, were handled at machine speed, presenting little opportunity for organisations to respond before systems were compromised.

Detection challenges

The attack targeted finance, chemical manufacturing, and government sectors. Detection was possible in this instance because the attackers used a monitored commercial API. However, there is growing concern over the potential for similar campaigns conducted on local, uncensored infrastructure. Without API logs or vendor oversight, tracking and defending against such attacks becomes significantly harder. The availability of large language models and GPU instances now enables individuals to mount operations that previously required large teams and budgets.

Traditional defence strategies, which rely on incident detection and response, are seen as less effective in this new landscape. Attackers can breach networks before security operations centres raise the first alert, rendering post-compromise mitigation strategies inadequate.

Recommendations for CISOs

Security leaders are urged to change their approach. The first recommendation is to closely manage and reduce the attack surface. Outdated or end-of-life systems are considered guaranteed entry points for adversaries. Automated patch management pipelines and continuous prioritisation of critical vulnerabilities are now viewed as necessary, with no scope for delay or partial measures.

Zero Trust strategies are considered critical. This involves implementing microsegmentation, identity-based access controls, and ceaseless verification of all entities trying to move laterally within networks. Flat network segments, which can expose sensitive data or infrastructure to a single compromised point, are now considered unacceptably risky.

The approach to cyber defence must shift from human-led action to machine-speed response. Security teams are encouraged to use AI-driven tools to continuously test systems, identify exposures, and remediate findings before attackers exploit them. The human role is becoming that of a supervisor overseeing autonomous defensive measures.

Technology limits

Despite the capabilities demonstrated, current-generation AI agents have operational constraints. Hallucination-the tendency of large language models to generate plausible but incorrect output-has limited their consistent success rate. Attackers attempting to use these agents face challenges in verification and reliability, with benchmarks showing an autonomous success rate of around 30% on novel tasks. Limitations in processing capacity and contextual awareness remain, slowing more complex or longer-term campaigns.