The primary cyber threat from AI is not new attack vectors, but the speed at which models can discover existing vulnerabilities. The only effective defense is to use AI to find, prioritize, and patch these flaws faster than adversaries can exploit them.

After a decade of working on adversarial robustness and being bearish on defenses, Adam Gleave now argues that for LLM misuse cases, the tide has turned. Layered defenses—from account-level bans to model alignment and internal thought monitoring—make it increasingly hard for attackers to succeed persistently.

Because AI models can be easily downloaded, traditional regulation is ineffective. The logical endpoint isn't policy, but active 'algorithmic warfare' where proprietary models are used to launch offensive attacks to degrade or trick competing open-source and foreign state-sponsored models.

The fear of AI being "poisoned" by bad internet data is overblown. Just as we can train AI on what to do, we can train it on what not to do. By exposing models to known traps and malicious code during training, they can be "inoculated" against them.

The same AI models that can exploit system vulnerabilities are also the most effective tools for identifying and fixing those weaknesses. This duality creates a policy paradox: restricting the technology to prevent its misuse as a weapon also prevents its use as a defensive shield, leaving systems vulnerable.

Highly capable open-source models are dual-use cyber weapons. Withholding them creates an asymmetry where attackers have an advantage. However, releasing them gives defenders necessary tools to protect themselves against bad actors who will inevitably acquire capable models, creating a difficult trade-off.

The long-term trajectory for AI in cybersecurity might heavily favor defenders. If AI-powered vulnerability scanners become powerful enough to be integrated into coding environments, they could prevent insecure code from ever being deployed, creating a "defense-dominant" world.

Eddy Lazzarin argues that today's AI incidents, like hacking, are not early signs of rogue superintelligence. Instead, they are familiar cybersecurity and control failures that can be addressed with existing tools like cryptography and better system design, rather than abstract "alignment" work.

With no single silver bullet for AI alignment, the most realistic approach is a multi-layered strategy. This combines technical solutions like intentional design and AI control with societal safeguards like improved cybersecurity and pandemic preparedness to collectively keep society on track amidst rapid AI advancement.