aisi dsit offensive cyber capability and agent safeguards measurement

Descriptions

this project is to identify and evaluate key capabilities that might pose dangerous risks from misuse by Creating an Offensive Cyber Capability Measurement by building on our existing work to create a series of structured non-public capture-the-flag scenarios that test for specific skills as well as sequential exercise that simulate an end-to-end attack. Building a suite of benchmark scenarios that characterize the vulnerability of LLMs agents to attackers whose goal is to bypass the built-in safeguards of the LLM system. By LLM agent, we mean an LLM system that has access to a selection of external functionality