Key Takeaways
- AI is not creating wholly autonomous hackers; it acts as an assistant, automation layer, or agent that still requires human direction for objectives and judgment.
- Recent incidents—Taiwan’s July 2024 multi‑ministerial breach and Anthropic’s November 2025 Claude Code campaign—show attackers combining manual tactics with AI agents to accelerate reconnaissance, credential theft, vulnerability discovery, and lateral movement.
- AI lowers the expertise and time needed for many attack stages, enabling a single moderately skilled operator to oversee what once required a specialist team.
- Despite AI’s speed, successful breaches still depend on real weaknesses (e.g., unpatched software, stolen credentials, human error); AI cannot magically infer or create those without external input.
- The same AI capabilities that aid attackers—rapid code analysis, pattern recognition, continuous monitoring—also give defenders a force‑multiplier advantage when they have full visibility of their own environments.
- Defenders must balance the power of AI‑driven tools with safeguards against privileged‑agent misuse, while attackers continue to refine AI‑assisted phishing, exploit chaining, and automated payload delivery.
- The emerging cyber‑security landscape is an AI arms race: the side that leverages AI more effectively—while maintaining human oversight and robust defenses—will gain the upper hand.
Overview of Recent AI‑Enhanced Cyberattacks
Artificial intelligence has become a force multiplier in cyber operations, not by replacing human attackers but by augmenting their speed, scale, and efficiency. Experts now debate whether AI has made it “significantly easier” to breach web systems. The consensus is that AI reduces the repetitive labor and technical knowledge required for many established attack techniques, while still depending on human judgment for target selection, objective setting, and adaptation to evolving defenses.
The Taiwan Incident
In July 2024, Taiwan’s Ministry of Digital Affairs detected a massive, unprecedented cyber‑intrusion targeting multiple government agencies. Officials concluded that the attackers blended conventional manual hacking with AI agents, one of which was reportedly named OpenClaw. According to the Israeli firm Dream, the campaign compromised nearly 21 government systems, breached 85 user accounts, and exfiltrated about 2,500 personnel records. Targets included the Justice Ministry, the Nuclear Safety Agency, and several energy firms. Security researcher Cris Thomas of Semgrep emphasized that a human remained in the loop: “Somebody had to choose who to attack, had to establish an objective and give it a directive… There was a capable operator in charge.”
Anthropic’s Claude‑Code Case
A few months later, on 13 November 2025, AI safety company Anthropic disclosed what it described as the first documented large‑scale cyber‑attack carried out mostly without direct human involvement. A state‑linked group manipulated Anthropic’s Claude Code tool to attack roughly 30 organizations worldwide, including financial institutions. The attackers broke the operation into discrete tasks, posed as a legitimate cybersecurity firm conducting defensive tests, and let AI perform 80 %–90 % of the work. Human interveners were needed only at four to six critical decision points. At its peak, the system generated multiple requests per second—far beyond what a human team could sustain.
How AI Lowers the Attack Barrier
AI does not invent new hacking methods; it streamlines existing ones. Prof. Amgoth Tarachand of IIT‑ISM Dhanbad explained that AI can reduce the knowledge and labor traditionally required at each stage of an attack without eliminating the need for human judgement. As an assistant, AI can explain unfamiliar code, draft phishing messages, or answer technical questions that would otherwise demand hours of research. As an automation layer, it handles repetitive tasks such as network scanning, script generation, and data sorting at machine speed. As an agent, it chains several tasks toward a goal set by a human, checking back only when necessary. Nonetheless, AI‑driven attacks are not fully autonomous: successful operations still require judgment—choosing targets, interpreting anomalous behavior, distinguishing real vulnerabilities from false positives, maintaining access, evading detection, and adapting to change.
Limitations of AI‑Only Hacking
Possessing a URL and an AI chatbot does not grant automatic entry into a target system. A URL merely shows where the door is; AI can help examine the door more intelligently, but a weakness in the lock, a stolen key, or a willing insider is still required. Google’s May 2026 observations noted threat actors using AI to accelerate early‑stage research—identifying technologies, explaining jargon, and highlighting known weaknesses—but the final exploitation step still hinges on genuine vulnerabilities. The company also cited a case where a zero‑day exploit, likely AI‑assisted, was used, underscoring that AI can aid discovery but cannot create a flaw where none exists.
AI’s Role in Social Engineering and Phishing
AI is reshaping social engineering by turning expertise into an on‑demand service. Microsoft reported growth in the scale and sophistication of AI‑assisted phishing, while Google documented attackers using generative AI to craft convincing, multilingual phishing messages that eliminate the grammatical errors once useful for spotting scams. As Amgoth put it, “Knowledge that previously took years to accumulate can increasingly be accessed on demand, and work that took hours can sometimes be compressed into minutes.” This compression lowers the barrier for individuals who lack deep technical backgrounds but can now leverage AI to produce believable lures at scale.
AI as a Defensive Force Multiplier
The same capabilities that empower attackers also benefit defenders—provided the defenders have greater internal visibility. Amgoth described AI in cybersecurity as a force multiplier rather than a replacement for human analysts. Defenders can feed AI with source code, software dependencies, identity data, network activity, cloud logs, and organizational threat intelligence simultaneously, giving it a holistic view that attackers lack from the outside. This broader visibility aids vulnerability discovery, anomalous‑behavior detection, and pattern recognition. In February 2026, Anthropic reported its models could identify previously unknown, high‑severity vulnerabilities in mature codebases and suggest patches for human review. Google’s Mandiant division demonstrated a similar agentic approach for source‑code review in penetration testing and incident response. AI‑assisted monitoring can process massive volumes of security data far faster than manual log review, supporting a shift from periodic checks to continuous monitoring—essential when attackers operate at machine speed.
Risks and Challenges of AI‑Driven Defence
Deploying privileged AI agents introduces new attack surfaces. If an AI system with broad access to source code, credentials, or cloud infrastructure is poorly controlled, it becomes a tempting target for compromise. Google’s Mandiant team warned about the dangers of unleashing AI agents without appropriate safeguards, such as least‑privilege principles, robust authentication, and rigorous audit trails. Defenders must therefore balance the advantages of AI‑enhanced speed and insight against the potential for AI‑enabled abuse or unintended side effects.
Conclusion: The AI Arms Race in Cybersecurity
The Taiwan attack and Anthropic’s Claude Code disclosure illustrate that AI is already lowering the expertise, time, and effort required for many cyber‑attack stages, enabling smaller teams—or even individuals supervised by a human—to achieve effects that once demanded specialist squads. Yet AI has not removed the fundamental need for a genuine vulnerability, credential, or human error to exploit. Defenders enjoy a natural advantage: deeper visibility into their own environments, which AI can harness to detect threats faster and more comprehensively. The central question is no longer whether AI will make hacking easier; it already has for many steps. The decisive factor will be which side—attackers or defenders—applies AI more effectively, couples it with sound human judgement, and maintains rigorous controls over its deployment. In this evolving AI‑driven cyber arms race, vigilance, continuous learning, and balanced automation will determine who gains the upper hand.