Security is not a property you establish once. Dependencies update, models are fine-tuned, prompts change and attackers adapt. RISE builds evaluation pipelines that run continuously and produce evidence an organisation can act on.
Components
- LLM-driven penetration testing, starting from PentestGPT and extending to evaluation of what makes an agent effective at real-world testing.
- Fuzzing and program analysis for memory-safety and protocol vulnerabilities, including LLM-guided greybox fuzzing.
- Safety and robustness benchmarks for language, vision-language and video models: jailbreak resistance, over-refusal, toxic-prompt detection, hallucination and harmful-content surfacing.
- Adversarial evaluation of learning systems against poisoning, backdoor and evasion attacks, building on the R layer’s defences.
- Assurance cases that combine these results into measurable, reproducible evidence.
The pipelines are designed to plug into development and operations workflows so that regressions are caught before deployment.