Mistral AI scientist Sagar Vaze points out a critical flaw in closed-model safety architectures: the guardrails are asymmetrically broken. Attackers can jailbreak capable models to probe for vulnerabilities, but when defenders try to run similar adversarial queries to stress-test their own systems, the same safety filters block them. The irony? Security researchers trying to harden infrastructure get flagged as threats, while actual bad actors find workarounds. Vaze's take: if you're serious about red-teaming at scale, you're forced to use open-weight models because closed APIs will refuse your legitimate security queries. This creates a perverse incentive structure where the people who need powerful models for defense are pushed toward open alternatives, while attackers get the benefit of closed-model capabilities through jailbreaks anyway. Classic case of safety theater backfiring on the good guys.