Key Takeaways
- LLMs have been documented generating functional exploit code, phishing campaigns, and malware variants in red team exercises
- Sonatype identified more than 454,600 new malicious packages in public software repositories in 2025
- Mandiant's M-Trends 2026 estimates the mean time to exploit at minus seven days, though it does not treat 2025 as the year breaches were the direct result of AI
- Three Japanese students aged 14 to 16 were arrested after using ChatGPT to improve a scheme targeting Rakuten Mobile accounts
- Defenders are patching at human speed; attackers are exploiting at machine speed — a fundamentally asymmetric race
The Facts
For decades, writing effective malware required deep technical skills: assembly language knowledge, operating system internals expertise, reverse engineering capability, and months of development time. The availability of frontier LLMs has lowered that barrier. Today, a motivated attacker with limited coding experience can get substantial help writing attack code.
The democratisation is documented at the extreme end of the capability spectrum. In a widely reported case in Japan, three students aged between 14 and 16 were arrested after using ChatGPT to improve the speed and efficiency of a scheme targeting Rakuten Mobile accounts; Japanese reports put the number of unauthorised logins at about 220,000. The case illustrates both the lowered barrier and the scale at which AI-assisted attacks can operate.
The software supply chain data tells the same story from the defender's perspective. Sonatype identified more than 454,600 new malicious packages in public repositories in 2025. How much of that growth is due to AI coding tools has not been established; the timing is consistent with, but does not prove, such a link.
Technical Deep-Dive
LLM-assisted malware development operates across several capability tiers. At the lowest sophistication, attackers use general-purpose chatbots to generate phishing email templates, social engineering scripts, and basic automation code. This tier requires no technical background and produces results that human-written equivalents could match — but at dramatically lower cost and with language quality that defeats simple text-based detection.
At medium sophistication, attackers use coding-focused AI assistants to generate proof-of-concept exploit code from CVE descriptions. Given a vulnerability description and the affected software's API structure, a current frontier model can generate plausible exploit code that requires testing and refinement but provides a substantial head start over manual development.
At the highest sophistication, Google's Threat Intelligence Group reported in January 2025 that government-backed hacking groups had used its Gemini model for reconnaissance on target organisations, research into vulnerabilities and payload development. Its conclusion was measured: "Rather than enabling disruptive change, generative AI allows threat actors to move faster and at higher volume."
Exploitation is also getting faster. VulnCheck found that 28.3% of vulnerabilities known to be exploited in Q1 2025 showed exploitation evidence within a day of CVE publication, and Mandiant's M-Trends 2026 estimates the mean time to exploit at minus seven days, meaning exploitation routinely comes before a patch. Mandiant does not attribute that to AI: it says it does not consider 2025 to be the year where breaches were the direct result of AI.
The ASEAN Perspective
ASEAN organisations are on the receiving end of this capability shift without the equivalent AI-powered defensive tools. Many enterprise security operations centres in the region still rely on signature-based detection, manual incident analysis, and periodic vulnerability scanning — defences built for human-speed attacks.
AI-powered defensive tools — AI-driven EDR, AI-assisted SOC automation, AI-accelerated threat hunting — are commercially available but require security engineering expertise to deploy and tune effectively. The talent constraint means that ASEAN enterprises are exposed to AI-speed attacks while their defences remain largely human-speed.
RECATOOLS Verdict
The LLM-assisted attack reality is not a future concern — it is the current operating environment. The response requires investment in AI-powered defensive tools to match AI-powered offensive capabilities, and realistic acknowledgement that the human-speed patch-and-respond cycle is no longer sufficient as a primary defence.
Sources
- Google Threat Intelligence Group, Adversarial Misuse of Generative AI (January 2025)
- Mandiant M-Trends 2026 Report
- VulnCheck, 2025 Q1 Trends in Vulnerability Exploitation
- Sonatype 2026 State of the Software Supply Chain
FAQ
Can AI really write malware? Yes — LLMs can generate functional exploit code, phishing templates, and malware variants. The skill barrier for basic attack tooling has been dramatically reduced.
What is the supply chain connection? Sonatype identified more than 454,600 new malicious packages in public repositories in 2025. Whether AI coding tools are driving that growth has not been established.
What happened with the three students and Rakuten? Three students in Japan aged 14 to 16 were arrested after using ChatGPT to improve a scheme targeting Rakuten Mobile accounts; Japanese reports put the number of unauthorised logins at about 220,000.
How fast are CVEs being exploited? VulnCheck found that 28.3% of vulnerabilities known to be exploited in Q1 2025 showed exploitation evidence within a day of CVE publication. Mandiant's M-Trends 2026 estimates the mean time to exploit at minus seven days, but does not attribute the speed-up to AI.
What defensive AI tools are available? AI-driven EDR (Endpoint Detection and Response), AI-assisted SIEM analysis, AI-powered vulnerability scanning, and AI-augmented threat hunting platforms are commercially available from major security vendors.
Correction, 18 October 2026: An earlier version attributed a figure on rapid exploitation to Mandiant; it comes from VulnCheck and refers to known exploited vulnerabilities. We have also corrected the description of the Rakuten Mobile case, removed references to two reports whose findings had been misdescribed, and removed an unsupported claim linking malicious package growth to specific tool releases.
View author profile → · Editorial policy
