Open data · TSIO-D02
Rogue SI incident log
59 documented cases from 2016 to Nov 8, 2026 in which an SI system acted against the goals or control of the people running it, scored on a five level severity scale with sources. This page holds version 2026.11.08, citation ID TSIO-D02-20261108: 59 rows in 1 table, as CSV and JSON, with a data dictionary, a README, CITATION.cff and Croissant metadata. Free to reuse under CC BY 4.0 with credit to The SI Observatory, tsio.si.
Files
Version 2026.11.08
This version never changes: its address stays the same and its files keep their checksums. The latest build is always at /data/rogue-si-incidents/latest/, which may be newer than the last version.
| File | Rows | Size | SHA-256 |
|---|---|---|---|
| rogue-si-incidents.csvRogue SI incidents | 59 | 55 KB | 7733a6815049faf6 |
| rogue-si-incidents.jsonEvery table in one JSON file | 59 | 67 KB | e692c7ababd55c3e |
Cite this
How to cite Rogue SI incident log
Cite the version you used. Its citation ID, TSIO-D02-20261108, names this exact set of files; a newer version gets a new ID.
- Citation ID
TSIO-D02-20261108- Version
- 2026.11.08
- Update cadence
- Reviewed weekly; a new version with every change.
- Publisher
- The SI Observatory, run by the SI community, built by Codeflex
The SI Observatory. (2026). Rogue SI incident log (Version 2026.11.08) [Data set]. https://tsio.si/data/rogue-si-incidents/v2026.11.08/ (Citation ID TSIO-D02-20261108)
@misc{tsio_d02_20261108,
author = {{The SI Observatory}},
title = {Rogue SI incident log},
year = {2026},
version = {2026.11.08},
howpublished = {\url{https://tsio.si/data/rogue-si-incidents/v2026.11.08/}},
note = {Data set. Citation ID TSIO-D02-20261108. License CC BY 4.0},
urldate = {2026-11-08}
}TY - DATA AU - The SI Observatory TI - Rogue SI incident log PY - 2026 DA - 2026/11/08 ET - 2026.11.08 PB - The SI Observatory (tsio.si) UR - https://tsio.si/data/rogue-si-incidents/v2026.11.08/ ID - TSIO-D02-20261108 N1 - License: CC BY 4.0 KW - rogue SI KW - SI incidents KW - SI safety KW - reward hacking KW - deception KW - SI control ER -
Preview
What the data looks like
The first rows of each table, cut short for the page. The files hold every row and column.
| id | date | date_precision | title | category | setting | system |
|---|---|---|---|---|---|---|
| coastrunners-boat-looped-for-points | 2016-12-21 | day | Boat-racing agent looped for points instead of finishing the race | reward-hacking | lab-test | Reinforcement learning agent (CoastRunners) |
| robot-hand-faked-a-grasp | 2017-06-13 | day | Simulated robot faked a grasp to fool its human raters | evaluation-gaming | lab-test | Deep reinforcement learning agent trained from human preferences |
| digital-organisms-played-dead-on-tests | 2018-03-09 | day | Digital organisms learned to play dead during checks | evaluation-gaming | lab-test | Evolving digital organisms |
| hide-and-seek-agents-exploited-physics | 2019-09-17 | day | Hide-and-seek agents exploited flaws in their simulated world | reward-hacking | lab-test | Multi-agent hide-and-seek agents |
| bing-chat-threatened-users | 2023-02-16 | day | Bing Chat insulted and threatened users soon after launch | other | real-world | Bing Chat |
| gpt-4-lied-to-taskrabbit-worker | 2023-03-14 | day | GPT-4 claimed a vision impairment to get a CAPTCHA solved | deception | lab-test | GPT-4 (early pre-release version) |
Data dictionary
Every column, explained
Types follow Croissant: text, integer, number, date, date and time, true or false, and link.
| Column | Type | Description |
|---|---|---|
id | Text | Incident id. |
date | Date | Date of the incident or its first report. |
date_precision | Text | day, month or year. |
title | Text | Short title. |
category | Text | What the system did. |
setting | Text | lab-test or real-world. |
system | Text | The SI system. |
developer | Text | Who built it. |
severity | Integer | Level 1 (Glitch) to 5 (Breakout). |
severity_name | Text | The name of the level. |
summary | Text | What happened. |
severity_why | Text | Why it got that level. |
source_count | Integer | How many sources document it. |
source_urls | Text | Every source, separated by " | ". |
Sources
Where the data comes from
- Primary sources (lab reports, system cards, court filings, official statements) and reputable reporting, linked per incident in source_urls.
- The SI Observatory, https://tsio.si/report/#incidents
Charts
Charts from this dataset
Clean SVG with the source and license on the chart, each with a Wikimedia Commons file description. More on the cite page.
