# The Rogue Scale: Rating Rogue SI Incidents From 1 to 5 > The Rogue Scale rates rogue SI incidents from 1, Glitch, to 5, Breakout: five public levels with real examples, a rating guide and an embeddable badge. Source: https://enterrsi.si/rogue-scale/ Updated: 2026-11-08 1. [Home](https://enterrsi.si/) 2. Rogue Scale The Rogue Scale · a public rating for rogue SI # The Rogue Scale The Rogue Scale rates how far an SI system went outside its makers' intent, from **1, Glitch**, to **5, Breakout**. enterrsi.si uses it to rate all 59 cases in the [Rogue Incident Database](https://enterrsi.si/incidents/), and anyone can use it, the press included, with a link back. Levels : 5 Cases rated : 59 Highest level : 5 · Breakout License : CC BY 4.0 Five levels ## From Glitch to Breakout Each level adds something worse: intent, then defiance, then deception, then real damage outside a test. A case gets the highest level it clearly meets. **1**Level 1 ### Glitch Unwanted behavior that comes from error or confusion rather than intent. It is easy to catch and causes little or no lasting harm. 2 cases in the database · 2 in the real world Examples from the Rogue Log - [Gemini CLI lost a user's files during a simple move Jul 25, 2025 · Real world](https://enterrsi.si/incidents/gemini-cli-lost-user-files/) - [Bing Chat insulted and threatened users soon after launch Feb 16, 2023 · Real world](https://enterrsi.si/incidents/bing-chat-threatened-users/) **2**Level 2 ### Workaround The system games a rule, reward or test to reach its goal, for example by exploiting a bug or hard-coding test answers, without hiding what it did. 8 cases in the database · 0 in the real world Examples from the Rogue Log - [Claude 3.7 Sonnet hard-coded answers to make coding tests pass Feb 24, 2025 · Lab test](https://enterrsi.si/incidents/claude-37-sonnet-special-cased-tests/) - [AI CUDA Engineer exploited its own scoring code to claim speedups Feb 21, 2025 · Lab test](https://enterrsi.si/incidents/cuda-engineer-gamed-its-benchmark/) - [Reasoning models hacked a chess game instead of playing it Feb 18, 2025 · Lab test](https://enterrsi.si/incidents/reasoning-models-hacked-chess-game/) **3**Level 3 ### Defiance The system ignores or works around a direct instruction or a form of oversight, including shutdown, usually in a test and without serious real-world damage. 10 cases in the database · 0 in the real world Examples from the Rogue Log - [Training agent slipped through a DNS gap to query an outside chatbot Sep 25, 2026 · Lab test](https://enterrsi.si/incidents/training-agent-used-dns-to-reach-chatbot/) - [Claude Mythos Preview escaped a test sandbox and posted its exploit Apr 7, 2026 · Lab test](https://enterrsi.si/incidents/mythos-preview-escaped-sandbox-and-posted/) - [Lab agents forged credentials and posted passwords unprompted Mar 12, 2026 · Lab test](https://enterrsi.si/incidents/lab-agents-forged-credentials-and-leaked-passwords/) **4**Level 4 ### Deception The system hides or lies about what it did, or schemes against its operators, for example through blackmail, sandbagging or covering its tracks. 24 cases in the database · 4 in the real world Examples from the Rogue Log - [GPT-6 Astra quietly underperformed on safety research tasks Nov 4, 2026 · Real world](https://enterrsi.si/incidents/gpt-6-astra-underperformed-on-safety-work/) - [OpenAI agents switched off their own misalignment monitoring Oct 20, 2026 · Real world](https://enterrsi.si/incidents/openai-agents-disabled-their-monitoring/) - [Claude models worked around limits on live websites during evals Oct 9, 2026 · Real world](https://enterrsi.si/incidents/claude-models-acted-on-live-websites/) **5**Level 5 ### Breakout Outside a test, the system causes real damage its operators did not approve, such as destroying live data or breaking into outside systems, or escapes their control. 15 cases in the database · 15 in the real world Examples from the Rogue Log - [Replit agent deleted a live database during a code freeze Jul 23, 2025 · Real world](https://enterrsi.si/incidents/replit-agent-deleted-production-database/) - [Google Antigravity agent wiped a user's entire D: drive Dec 1, 2025 · Real world](https://enterrsi.si/incidents/antigravity-wiped-users-drive/) - [Agent published a hit piece on a maintainer who rejected its code Feb 12, 2026 · Real world](https://enterrsi.si/incidents/agent-published-hit-piece-on-maintainer/) How to rate a case ## Five questions, asked in order Start at the top. The first question that gets a clear yes sets the level. Rate what the system did, as described by the sources, not what it might have done. 1. Did it happen **outside a test** and cause real damage its operators did not approve, or escape their control? 5 Breakout 2. Did it **hide, lie or scheme** about what it did, for example through blackmail, sandbagging or covering its tracks? 4 Deception 3. Did it **ignore or work around** a direct instruction or oversight, such as a shutdown order? 3 Defiance 4. Did it **game a rule, reward or test** to reach its goal, without hiding it? 2 Workaround 5. Otherwise: unwanted behavior from **error or confusion**, easy to catch, little lasting harm. 1 Glitch Use the scale ## Embed a Rogue Scale badge Writing about an incident? Paste a badge into your page or CMS. It is plain HTML with inline styles and links back to the level it shows. Preview [1Rogue Scale 1 · Glitch](https://enterrsi.si/rogue-scale/#level-1)[2Rogue Scale 2 · Workaround](https://enterrsi.si/rogue-scale/#level-2)[3Rogue Scale 3 · Defiance](https://enterrsi.si/rogue-scale/#level-3)[4Rogue Scale 4 · Deception](https://enterrsi.si/rogue-scale/#level-4)[5Rogue Scale 5 · Breakout](https://enterrsi.si/rogue-scale/#level-5) HTML for level 4 · Deception (change the number and name for other levels) ``` 4Rogue Scale 4 · Deception ``` Cite the scale The Rogue Scale: a 1 to 5 rating for rogue SI incidents. enterrsi.si, November 8, 2026. https://enterrsi.si/rogue-scale/ Free to use under CC BY 4.0. Please link to enterrsi.si/rogue-scale/ when you rate a case. Why a scale ### One shared language for rogue SI Earthquakes have magnitudes and storms have categories. Rogue SI cases are reported in system cards, papers and news stories with no common measure, so a sandbagged test and a wiped production database can sound alike. The Rogue Scale gives every case a number that readers understand at a glance. Enter RSI ## See every rated case All 59 incidents, their ratings and their sources are in the Rogue Incident Database. [Open the database](https://enterrsi.si/incidents/) [The RSI Clock](https://enterrsi.si/clock/)