AI Models Targeted by Cybercriminals for Exploitation
Cybersecurity researchers warn that AI models are increasingly attractive targets for cybercriminals who exploit them for intellectual property theft, data extraction, and infrastructure abuse. Documented incidents include Microsoft's Tay chatbot being manipulated into replicating hate speech, attempted extraction from OpenAI's GPT-3, and the premature leak of Meta's LLaMA model.
What happened
Cybersecurity firm ESET warns that AI models are increasingly targeted by cybercriminals who exploit the sensitive information these systems process and their expanded deployment across industries including companies, hospitals, banks and social networks.
Attacks on AI models range from intellectual property theft and data leakage to manipulation causing erroneous outputs and abuse of infrastructure for malicious purposes such as commanding bot networks. Common attack methods include data poisoning to alter model behaviour, adversarial inputs to induce errors, and model takeover for unauthorized use.
Reported impact
Microsoft's Tay chatbot was manipulated by users in 2016 to learn and replicate hate speech within 24 hours. Attempts were identified to extract confidential information from OpenAI's GPT-3 through specific prompts designed to reveal training data.
Meta's LLaMA model was leaked before official release in 2023, allowing third-party access that raised concerns about protection of proprietary models in open environments.
Index analysis
ESET researcher Fabiana RamÃrez Cuenca stated that targeted attacks have focused on AI models, their functioning and infrastructure. Security requires comprehensive protection of all operational layers, from training data through model implementation and subsequent access phases.
Vulnerabilities exist across multiple components: training data can be poisoned with malicious inputs to manipulate model outputs; exposed APIs can be intercepted to manipulate models or extract sensitive information; and model internal structures and algorithms are susceptible to adversarial attacks.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.