The safeguard that cannot tell a cure from a weapon

On September 10, 2026, the AI company Anthropic published a report on people trying to bend its models toward harm. Over nine months it had disrupted roughly 35 concerning efforts, five of them laid out in detail. In May, one user asked for help drafting a grant to make chikungunya more lethal by passing it through live animals again and again; the work was headed for a military institute. Another traded thousands of messages working out how to give avian influenza the airborne reach it does not yet have.

The report's most honest sentence is also its most defeated. The safeguards, it concedes, cannot tell a virus researcher from a weapons developer. The question that designs a vaccine designs the plague. A biodefense scholar, Gregory Koblentz, calls such knowledge multi-use: the same facts serve cure, defense, and slaughter, and nothing in the facts decides which.

What decides is the heart, and the heart is the one thing no filter can read.

This is the wall the prophet reached three thousand years before anyone wrote a content policy. He was not talking about software. He was naming something permanent in us: intent hides inside the same words, and who can know it? The screen fails at the exact seam he said every human judgment fails on. It reads the sentence and cannot touch the will beneath it.

His next line refuses to let the question hang.

Jeremiah 17:10I, Yahweh, search the mind, I try the heart, even to give every man according to his ways, according to the fruit of his doings.

What no safeguard can screen is still not hidden. The heart a model logs as one more prompt is already searched and weighed, and given back exactly what it earns. The danger in the report was never in the knowledge. It was always in the person holding it. And the intent no filter can read is read completely by God.