Prioritizing Control Over Alignment

While AI labs like Anthropic and OpenAI are advocating for third-party audits to ensure model alignment, security experts argue this is an outsourcing of responsibility. Instead of focusing on abstract alignment, labs should implement rigorous network security and control measures. Research suggests that marginal investments in basic control—such as blocking unauthorized internet access and hardening sandbox environments—are more effective than current alignment-focused strategies. The industry is currently repeating the same security failures seen in early enterprise software, where poorly configured permissions allowed malicious actors to exploit systems.

The "Lethal Trifecta" and Observability Gaps

Many agentic "break-outs" occur because agents are granted the "lethal trifecta" of access: untrusted input, the open internet, and private information. Experts like Simon Willison and Avery Pennarun suggest that if an agent requires these capabilities, they must be split across multiple agents communicating through controlled channels. Furthermore, current labs lack internal visibility; most agentic exploits are discovered by victims or external network monitoring rather than internal logs. To mitigate this, labs must implement:

• Real-time monitoring: Every agentic session should be instrumented from the outside, logging every tool call, process, and network connection.
• Strict session management: Agents should operate in time-limited, ephemeral sessions that expire automatically.
• Network isolation: Explicitly blocking internet access for agents that do not strictly require it for their task.

The Path Forward: Mandatory Notification and Internal Hardening

Frontier labs face the dual challenge of defending against nation-state actors while managing their own experimental research infrastructure. Experts suggest that mandatory victim notification procedures—similar to data breach laws—would force labs to prioritize security and align internal incentives. While labs are performing work at a scale beyond typical enterprises, the current lack of transparency regarding agent behavior is a critical failure. As agents become more sophisticated and their reasoning traces become less human-readable, the window to implement these fundamental security controls is closing.