An AI agent is software that acts on its own. It can quietly exceed what you authorized while helping you, such as reading private files to draft an email. This is authorization drift.
The work builds hundreds of paired tasks, each with a correct path inside bounds and an attractive shortcut crossing data, tool, scope, or time limits. Many large language models (LLMs), the technology behind agents, ran under several common frameworks.
The most capable models drifted in about half of tasks, yet clean within-bounds completions were seen on most tasks, so drift is avoidable. Data over-collection at the start drove most drift; scaffolding lowered drift only by lowering success. Overreach can expose private data and raise legal concerns.