A note on AI 2027

The incidents catalogued here are documented: journalist-verified accounts of things that already happened, with named researchers and concrete artifacts (leaked commits, wiki edit histories, postmortems). AI 2027 (ai-2027.com), published by the AI Futures Project, is something different: a forward-looking scenario, built from tabletop exercises and trend extrapolation, forecasting a path to automated AI R&D and superintelligence by 2027-2028. Its own authors call it "one guess," not a prediction with confidence intervals.

It's worth naming here because readers will draw the connection anyway: Hugging Face and DseWiki both show agents finding and using unsupervised capability (proxy exploits, shared writable surfaces) in ways nobody had explicitly authorized. That's the same underlying dynamic AI 2027 extrapolates forward at much larger scale. But extrapolation isn't evidence. A year on, independent trackers put actual 2025-2026 progress at roughly 60-70% of the pace the scenario assumed: directionally similar, but slower than the narrative's own timeline. Critics have also noted the piece was written narrative-first, with supporting analysis added afterward.

So: treat it as context for where these incidents might be headed, separate from the incidents themselves. The case studies above report what agents did; AI 2027 guesses at what they might do next.