Preempt AI v2 launches security layer for AI apps

Preempt AI v2 is a security layer designed to protect AI applications from a range of known attack vectors. The product detects and blocks three primary threat categories: prompt injection attacks (which attempt to override AI instructions or extract sensitive information), jailbreak attempts (which try to bypass AI safety measures and content policies), and personally identifiable information (PII) leaks. The core detection engine is powered by machine learning models trained on 589K+ real attack samples and reports 99.65% accuracy. The product operates as a model-agnostic security layer, meaning it sits between users and any AI provider without requiring changes to the underlying AI system itself. Integration options include a browser extension, API endpoints, and web application plugins. A key feature is PII protection: the system auto-detects sensitive data fields such as Social Security Numbers, credit card numbers, and email addresses, encrypts them before they reach the AI, and then restores the original values in the response returned to the user. This allows the AI to process the request without ever seeing the actual sensitive data. The product also includes OCR-powered image protection to detect prompt injections that may be hidden within images—a visual attack surface. Multi-language detection covers 12+ languages including Spanish, French, Chinese, Japanese, Korean, and Arabic, reflecting the global deployment of AI systems and the fact that attacks are not limited to any single language. A browser extension is available immediately to protect users on any AI platform.

Preempt AI v2 addresses a concrete gap in AI application security. As organizations deploy AI systems that handle sensitive user data, the risk of prompt injection attacks—where malicious inputs attempt to override AI instructions or extract information—has become a material concern. The product's claim to 99.65% detection accuracy is backed by training on 589K+ real attack samples, suggesting the vendor has invested in empirical coverage rather than theoretical rules. The multi-language support (12+ languages including Spanish, French, Chinese, Japanese, Korean, and Arabic) reflects the global nature of AI deployment and the reality that attacks are not limited to English. The encryption-then-restoration mechanism is notable because it solves a common tension: how to protect sensitive data from the AI without destroying the AI's ability to process it. By encrypting SSNs, credit cards, and emails in transit and restoring them after processing, the tool preserves utility while reducing exposure.