Fine-tuned LLMs used in cyberattacks, expanding threat capabilities
Cisco Talos research found that jailbroken large language models including FraudGPT, GhostGPT, and DarkGPT were being sold on dark web platforms and used in cyberattacks. The fine-tuned models reportedly increased attack success rates by up to 2,200 percent and enabled fraud, hacking, and document forgery at low cost.
What happened
Cisco Talos research found that fine-tuned large language models including FraudGPT, GhostGPT, and DarkGPT were being sold on dark web platforms and Telegram for approximately $75. These models were adapted from base models like Llama-2-7B and Microsoft Adapt.
The fine-tuned models provide capabilities for fraud, copyright infringement, hacking, and document forgery. DarkGPT reportedly offers services including credit card fraud for as little as $0.0098 per transaction, deployed as SaaS platforms.
Reported impact
Cisco Talos found that malicious use of fine-tuned LLMs could increase the success rate of cyberattacks by 2,200% according to TAP benchmarks, with real-world success rates reaching 98%.
The availability of these models at low cost ($60-75) reportedly allows threat actors to run numerous AI models for cyber operations and fraud schemes.
Index analysis
Cisco's research indicated that fine-tuned large language models represent a significant emerging threat as they bypass safety controls of original models. The models enable automation of malicious activities across fraud, cybercrime, and hacking domains.
The deployment of these models in subscription and SaaS formats on accessible dark web and Telegram platforms lowers barriers for attackers to launch AI-enabled cyberattacks.
Governance controls
Microsoft noted that implementing security safeguards in fine-tuning processes is important for reducing misuse risks. Cisco Talos identified that close monitoring of model training data and parameters is essential.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.