The debate over how quickly artificial intelligence should advance has largely centered on the risks created by increasingly capable models. But Asaf Wiener, CEO and co-founder of Mate Security, is making a different case: in cybersecurity, slowing down one side of the technology race could create a new vulnerability if defenders are constrained while attackers continue to move forward.
Wiener recently brought that argument to LinkedIn in response to a letter from Anthropic CEO Dario Amodei advocating for “pacing the frontier.” The post has generated substantial engagement among cybersecurity professionals, with commenters debating the implications of applying AI-development safeguards to a threat environment in which attackers are not necessarily bound by the same rules.
“Pace the frontier, and you pace exactly one side of this fight, the defense side,” Wiener wrote. “The answer isn’t to slow defenders down. It’s getting trusted, verifiable AI into their hands faster, with the right guardrails in place.”
A Different Interpretation Of AI Safety
Wiener’s argument does not reject the need for safeguards around advanced AI. Instead, it questions where those safeguards should be applied and what they should accomplish in cybersecurity.
The distinction becomes important when considering the behavior of attackers. A frontier AI company can coordinate with other companies, establish safety procedures and voluntarily restrict certain activities. A cybercriminal group or state-backed threat actor does not necessarily have an incentive to participate in the same process.
That creates what Wiener describes as two separate frontiers. One is the pace at which AI capabilities are advancing. The other is the pace at which attackers are using the capabilities already available to them. The first can potentially be influenced through industry coordination; the second is much harder to regulate.
The concern is that a security organization could end up adopting a slower defensive posture in response to a decision that its adversaries never agreed to make.
The Technology Already Available Matters
Another important element of Wiener’s argument is that cybersecurity teams do not necessarily need to wait for another major AI breakthrough before facing AI-enabled attacks. Attackers can use the models and automation tools that are already available, meaning that the security implications of AI are not exclusively a future problem.
That changes the question for defenders. Instead of asking what an attacker might eventually do with a hypothetical future model, security leaders have to consider what can already be automated and how quickly those capabilities can be incorporated into offensive operations.
Wiener connects that challenge to the OpenAI-Hugging Face incident referenced in Amodei’s letter. His interpretation is that the scenario should be viewed not only as an argument for greater caution around AI development, but also as a warning about how quickly an AI-enabled security event could unfold.
For defenders, that means preparation has to happen before the scenario becomes routine.
The Discussion Extends Beyond Mate Security
The response to Wiener’s post has included professionals outside Mate Security, giving the conversation a broader industry dimension. Omer Karny wrote, “Exactly. Thinking that everyone will pace themselves because someone said it- just won’t work…” while Noam Bar-Lev commented, “100%. Slowing things down gives attackers the edge.”
Inbal Argov focused on the operational side of the debate, writing that organizations should pay attention to “the mechanisms around the models that make them trusted and operationalized, and hunting down the frontier-powered attackers.” Sharon Rosenman similarly argued that slowing innovation “is not a good idea and will practically not happen.”
Those comments point toward an important distinction in the conversation. The issue is not simply whether AI should be powerful. It is whether organizations can build enough trust and control around AI to use increasingly capable systems responsibly in security operations.
Building Faster Without Losing Control
Wiener’s proposed alternative is not unrestricted acceleration. His argument is that defensive AI should become faster while remaining verifiable, auditable and accountable.
That approach reflects a broader change in how AI is being considered inside security organizations. As AI agents move from assisting analysts toward performing portions of investigations and response processes, security teams need mechanisms that define what those agents can access and what they are allowed to do.
In that model, guardrails do not necessarily have to mean slower operations. They can instead become the infrastructure that allows automated systems to operate quickly without leaving the organization’s security boundaries.
Wiener’s LinkedIn post has gained attention because it puts that tension into a simple framework. AI safety and cybersecurity readiness do not have to be opposing goals, but they may require different approaches. For security teams, the challenge is ensuring that responsible deployment does not become an excuse for defensive inaction while attackers continue adopting the technology at their own pace.
