Positive Technologies has introduced the LLM Firewall module within PT Application Firewall to protect corporate AI services from attacks targeting large language models.
Modern organizations increasingly entrust AI assistants with sensitive documents, customer data, and system access, making it critical to prevent unauthorized control by external actors. The new module analyzes prompts sent to language models and the corresponding outputs, searching for attempts to bypass built-in restrictions, extract system instructions, disclose confidential information, or redirect agent functions for malicious purposes.
One primary focus is countering prompt injection attacks. Malicious instructions can be delivered directly to the bot or concealed within documents that the model processes. While appearing as ordinary text, these inputs aim to appoint the neural network a new controller. The module can either block suspicious interactions or log them for further investigation.
The solution also detects personal data, passwords, access keys, and toxic or prohibited content. Inspection operates bidirectionally, monitoring both inputs to the model and outputs returned to users. This helps prevent leaks caused not only by sophisticated attackers but also by overly cooperative employees.
LLM Firewall is already integrated into PT AF Cloud and connects to application-model exchanges through APIs compatible with OpenAI Chat Completions and Responses. No separate management console is required. The module is scheduled for addition to PT AF PRO for on-premises deployments at a later date.
Administrators can define distinct policies for various AI services, select monitoring or blocking actions, and configure request and response checks independently. All events are forwarded to PT Application Firewall for analysis and incident investigation.
