New AI security research with clear technical context
3 results
This paper identifies a new attack vector scenario called 'Denial of Deadline', where targeted workload bursts push the latency limit for the slow, high-accuracy inference path beyond its deadline. This causes high-quality predictions to be discarded and drastically reduces overall system utility.
Security agent evaluations are often incomplete because they only measure peak offensive capability under generous inference budgets. In operational security, every reasoning step, tool call, or telemetry query consumes budget that limits actual success rates.
This paper argues that traditional penetration testing is insufficient for AI systems, as adversaries can alter system behavior through prompt injection or data poisoning without compromising the underlying infrastructure. It proposes a new framework focused on violations of operational behavioral objectives.