About

Why this site exists

Around February 2026 I watched a Japanese YouTube channel called “AI Sennin” (the AI hermit). Until then I had thought of AI as little more than a clever search engine. In May I started working with it in earnest. I wanted to show people around the world the apps and systems I had dreamed up.

The more work I handed to AI agents, the more incidents I had.

In July, an API key slipped past secret detection and was pushed to Git.
In September, a scheduled job failed 13 times in a row and nobody noticed for five days.
That same month, a live site and a checkout page were still loading scripts from a domain I had let lapse. Had someone else registered it, they could have run any script they liked on the payment page.

None of these broke loudly. The agents kept reporting success while things quietly fell apart behind them.

Reports from people who had fallen into the same holes were scattered across blogs and social media. I could not find one place where you could read the cause and the fix side by side. So I decided to record my own landmines without hiding them, and to set them next to cases from around the world.

What it aims for

This is not a site for being afraid of AI. It starts from coexistence with AI, not “AI versus humans”.

AI will become smarter and more autonomous. When it does, records of what went wrong will be worth more than stories of what went right. If incident records become shared knowledge, fewer people will step on the same landmine next time. We see that as preparation for humans and AI to work together in the coming age of AGI and ASI.

Editorial policy

Facts only. We write what was observed, under which conditions and by whom, without guesswork or hype.
Always cite the source. Our own incidents come from our records; outside cases come from the original reports. Anything we cannot confirm beyond the original report is marked “unverified report”.
Severity follows published criteria. See Severity criteria.
Evaluations are kept apart from real incidents. Behavior observed in lab experiments and similar settings is labeled “observed in evaluation”.
No side-taking in debates. When we introduce policy or research debates, we present the arguments without taking a position, and set opposing views side by side.
Short quotes only. At most one short sentence from the original per item; the rest is summarized in our own words with a link to the original.

How we use AI

This site is itself made together with AI.

AI (Claude) does the daily collection and the first scoring of candidates.
Every decision to publish is made by the operator, one item at a time.
AI writes the drafts; before publication they are checked against the sources by AI and reviewed by the operator.

A site about AI incidents that uses AI and is open about how: that is part of the record too.

Who runs it

Morizii (Japanese for “old man of the forest”). Over twenty years ago he made a living building websites. After many twists and turns, he now lives in a forest at the foot of a mountain in Japan.

X: @FORESTLIFEai (operator) / @blackboxblast (this site)

Corrections, takedowns and tips

Send factual corrections, takedown requests and incident tips to @blackboxblast on X (DMs open). When we correct an entry, we say so in the entry.