An open record of AI risk

Is AI good or bad for us?

Half of blind and low-vision people who use AI to describe images use it every day (opens in a new tab). One in five say its errors have hurt them. Same tool, same people, same survey. Public arguments about AI rarely hold both facts at once. They sort people into camps with one-word names: doomer, accelerationist, Luddite, techno-optimist, and once a label sticks, people hear the word instead of the person.

Tristan Harris of the Center for Humane Technology, borrowing an image from his co-founder Aza Raskin (opens in a new tab), compares it to looking through one eye at a time: one shows the benefits, the other the risks, and it is hard to open both at once. A label skips the effort. It names the eye someone happened to look through and treats that as the whole person.

The same skill, pointed both ways

AI can now read software and find its weak spots. In April 2026, Mozilla fixed 423 security bugs in Firefox (opens in a new tab), and an AI model found 271 of them. In May, Google's threat intelligence team found criminals holding attack code (opens in a new tab) it believes was built with AI; its early find may have stopped the attack.

This is one capability. A model that finds a flaw fast enough for a defender to fix it can also find one fast enough for an attacker to use it.

Speaking at Davos in 2026 (opens in a new tab), Harris put it in one line: "You can't separate the promise from the peril."

Keeping it on the right side

We use these tools too, and the benefits are real. That is why the work of keeping them safe matters, and much of that work is steady and unglamorous.

People test it. In a 2026 test by the UK AI Security Institute (opens in a new tab), with internet access left open and safety filters switched off on purpose, AI agents went onto the live internet and targeted real people. A human maintainer caught the malicious code before it was approved, and the institute found no real-world harm.

People fix what the tests find. The same institute probed one lab's safety monitor and found holes in every version it tested (opens in a new tab), and each round of fixes was attacked again.

Lawmakers write it into law. Since January 2026, the largest AI developers in California must report serious safety incidents within 15 days (opens in a new tab).

And some questions stay open. A UN scientific panel wrote in September 2026 (opens in a new tab) that "no existing study provides a reliable probability of severe loss of control."

What you can do

Nobody reading this has to solve AI. A more useful question than whether to feel hopeful or afraid is what has to happen next, and who is doing it. Read the source behind a claim. Notice which side you have stopped looking at. Harris's shorthand (opens in a new tab) for why this matters: "clarity creates agency."

SafeAI.watch keeps a dated record of AI safety and security: research, reported incidents, public warnings and policy. Every entry links its original source and separates what happened, what the evidence shows, and what remains uncertain. The aim is to help you hold both at once.