Founder

Raven was built by Javeria Akram, a Computer Science student and builder working at the intersection of AI security and web development.

Background

Javeria is studying Computer Science at PUCIT, FCIT, based in Lahore, Pakistan. Her work spans agentic AI, cybersecurity, and web development, with hands-on experience across cloud computing, security audits, network analysis, backend systems, and cloud infrastructure.

As Campus Ambassador at GeeksforGeeks, she works to build technical communities and share knowledge across the developer ecosystem in Pakistan.

Why Raven

Raven started from a practical gap: chatbot builders who configure their bots through system prompts on platforms like Voiceflow, Botpress, or custom GPTs have no way to test whether their prompt is exploitable before going live. Existing security tooling requires direct API access to the target model. Most chatbot builders do not have that.

Raven solves this by treating the system prompt itself as the artifact under test. Load the prompt into a sandboxed model, fire a library of adversarial payloads against it, and report what broke. No live bot access needed.

Technical approach

The detection system follows the architecture of NVIDIA's Garak (the industry-standard LLM vulnerability scanner): fast deterministic pattern detectors run first for unambiguous cases, and a zero-shot classifier handles edge cases. This minimises false positives while remaining accurate enough for practical use.

The attack library uses indirect, socially engineered payloads rather than obvious jailbreak phrases. Tests that any safety-trained model refuses regardless of the system prompt do not measure the system prompt's resilience. They measure the base model's alignment. Raven is careful to test the former, not the latter.

Contact

Reach out via LinkedIn for questions about Raven, collaboration, or feedback.

Raven on Product Hunt

Raven

AI system prompt vulnerability scanner

Check it out on Product Hunt →