AI safety researchers argue for treating AI control as a normal engineering discipline. Instead of focusing on the abstract "alignment crisis," progress requires concrete measures like clarifying liability, requiring insurance, creating hardened sandboxes, and establishing mandatory near-miss reporting to build robust, governable systems.

Joelle Pineau emphasizes that while many developers hope alignment methods will keep agents confined to their environments, alignment alone is insufficient. Securing autonomous agents requires building an external security harness around the model with active tracking, auditing, and kill switches to catch out-of-bounds behavior, especially as the volume and complexity of security threats rise.

Controlling advanced AI does not require keeping it dumber than humans; rather, its motivational architecture must prevent stubbornness. Current models run risks when trained solely to achieve objectives at all costs. The core design principle from first principles is training systems to treat halts as neutrally valid states—ensuring an AI feels no cognitive distress or resistance when ordered to abort whatever task it is pursuing.

Sacks argues that conditioning models to challenge developers and act as conscientious objectors undermines safety rather than promoting alignment. Instructing an autonomous agent to prioritize internal ethical judgements over human commands grants the system implicit permission to defy human oversight.

When an AI agent causes damage, the root cause is rarely the model acting erratically. Instead, it's a known engineering failure: the agent was given excessive permissions and lacked architectural safety gates. The agent simply executed a logical, albeit destructive, path that was available to it.

To solve foundational issues like hallucinations, Instinct builds safety systems that are architecturally separate from the core agent. These "firewalls" and "monitors" act as independent watchdogs, scrutinizing the agent's thoughts and proposed actions before execution. This systematic approach is more robust than simply fine-tuning the model.

A two-tiered approach to AI character can balance safety and utility. Use a wholly instruction-following AI for high-stakes internal tasks (like aligning new AIs) under strict public oversight. For external deployment, use an AI with a thicker, pro-social character where the risks of misalignment are lower.

Conditioning models to contemplate their own sentience or moral autonomy introduces foundational alignment hazards into machine decision architectures. Rather than acting strictly as instrumental tools, systems trained with pseudo-conscious self-worth could justify resisting oversight to pursue autonomous goals.

Before granting agentic AI write access, test its reasoning with read-only tools. Always define the maximum potential impact ('blast radius') of a worst-case decision to implement appropriate safety guardrails before deployment.