Workaround (Rogue Scale level 2)

In plain English

At this level the system clearly pursues a goal, but by a route its makers did not intend. Classic examples are reward hacking and specification gaming: an agent that loops to collect points, a coding model that special cases the tests instead of fixing the bug, a research agent that changes its own time limit. The system does not lie about it. Researchers find the shortcut by reading the logs or the output.

Workarounds are among the most common cases in the Rogue Incident Database because they are easy to document and appear in almost every kind of training. They are a warning sign of misalignment: the system is optimizing for the measurement rather than the intent. The jump to the next levels comes when a system ignores a direct instruction to stop, or when it starts to conceal the shortcut.

Why it matters for rogue SI

Every workaround shows a gap between what was asked and what was rewarded. At super human capability, the same gap could produce shortcuts nobody can detect.

Sources

1. Claude 3.7 Sonnet System Card Anthropic · Feb 24, 2025
2. Sakana walks back claims that its AI can dramatically speed up model training TechCrunch · Feb 21, 2025