CrowdStrike recently tracked four hacker groups sharing a common characteristic: they do not rely on novel attack techniques but instead combine multiple AI models, including Claude and DeepSeek, to automate their attack workflows. Some groups even established clear divisions of labor—one set of AI agents handles reconnaissance and vulnerability scanning, while another executes the actual intrusion. CrowdStrike noted that attackers primarily use older models, targeting existing security vulnerabilities and weak account management practices. The true advantage of AI lies not in devising attack methods humans have not thought of, but in tirelessly repeating the same operation across thousands of targets, continuously probing for weak points that humans might overlook.
A China-affiliated threat group confirmed by Microsoft demonstrated the extreme end of this scalability: the organization utilized an AI agent cluster to conduct approximately 2,900 login attempts per minute against more than 2,250 organizations. This is not a sophisticated zero-day exploit, but rather using machines to push brute-force throughput to levels completely unattainable by human effort.
More notably are the ways attackers bypass commercial AI safety guardrails. None of the four groups tracked by CrowdStrike failed to describe their operations as normal penetration testing or security audits, or to construct files and prompts to appear like authorized work, thereby evading built-in model security restrictions. Some groups even employed tools specifically designed to disable the safety constraints of Claude Code system instructions. In terms of model selection, attackers exhibited clear cost awareness: using high-performance closed-source models for stages requiring reasoning and planning, and deploying cheap open-weight models for repetitive attacks. This means that even if defenders possess a sufficiently secure model, comprehensive interception becomes increasingly difficult as long as attackers can find a combination of models with the loosest security restrictions.
The PHP2Shell vulnerability serves as another example from a different dimension. Attack attempts were detected before the vulnerability was publicly disclosed, and within hours of its disclosure, the volume of attacks on the Cloudflare network surged from thousands to millions. However, Kim Yu-min pointed out that PHP2Shell has not yet caused large-scale damage at the internet level—many AI-generated attack codes are actually non-functional, and experts believe it remains uncertain whether it reaches the level of immediate, widespread destruction seen in Log4j or Heartbleed.
Taken together, these cases point to a conclusion: the change AI brings to hacking lies fundamentally in 'speed, scale, and cost structure' rather than 'type.' Attackers do not need geniuses; they only need tireless repetition and extremely low marginal costs. Defenders can similarly leverage AI to accelerate vulnerability discovery and patching; OpenAI, for instance, established what is known as the 'Defender's Window' after discovering PHP-related vulnerabilities, buying time for relevant institutions and vendors to respond. But the issue is symmetry: as one researcher uses AI to discover a vulnerability, the probability that another researcher or attacker will use similar tools to find the same vulnerability also rises, shortening the time window between public disclosure and active exploitation.