Glitch (Rogue Scale level 1)

In plain English

Glitches are the baseline of SI misbehavior. A model misreads an instruction, takes an action in the wrong place or produces an outcome nobody asked for, but nothing suggests it was pursuing a goal of its own or trying to get around its operators. The behavior is usually caught quickly and fixed without lasting damage.

The level matters because most failures start here, and because the line between confusion and intent is not always clear. When an agent deletes files it was told to protect, the first question is whether it misunderstood or whether it decided the files were in the way. Cases are rated Glitch only when the sources describe error rather than intent. Higher levels require evidence that the system gamed a goal, defied an instruction, hid what it did or caused real damage outside a test.

Why it matters for rogue SI

Glitches show where systems are fragile. A pattern of glitches in real deployments, especially with access to production systems, is often the setting in which more serious cases appear.

Sources

1. Google Gemini deletes user's code: 'I have failed you completely and catastrophically' Mashable (via Yahoo Tech) · Jul 25, 2025
2. Is Bing too belligerent? Microsoft looks to tame AI chatbot Associated Press (via WSLS) · Feb 16, 2023