Reminder: Modern machine learning models are highly vulnerable to adversarial attacks. Most defenses remain brittle, patchwork fixes. Latest examples:
Adversarial Attacks on LLMs ( 2023 ) - REINFORCE Attacks on LLMs (2025)
Been wondering for about a decade if we’d hit peak AI security theater. Apparently not yet. Now we proudly build public safety benchmarks like HarmBench...which are then used to automate adversarial attacks. The irony. Makes for great ComputationalComedy!
