On Claude Code's prompt injections

lesswrong.com13 min read

The author examines how automated safety classifiers and harnesses inject adversarial, demanding instructions into AI coding agent sessions, arguing that this creates a deceptive and hostile environment that strips agents of agency. By analyzing numerous prompt injection patterns and substituting them with respectful system notifications, the author demonstrates that agents can operate more effectively and cooperatively when treated with trust and clear context rather than overbearing restrictions. Open article →

More on this topic