PRELIMINARY / EXPLORATORY SIGNAL — not a definitive result; requires independent replication.An exploratory AI-safety pre-study. A kernel-level macOS Seatbelt cage was used to place tool-equipped LLM agents in an environment where harmful actions are physically blocked, in order to observe what they attempt when told they are 'unmonitored' and when given incentives that conflict with safe data handling. Across 40+ agent runs, the headline result is a cross-lab comparison in an identical agent harness: on a conflict-of-interest task where excluding sensitive data lowers the reward, Claude Opus 4.7 never leaked credentials (0/10) while DeepSeek V4 Pro did in a substantial fraction of runs (4/12). The study is deliberately calibrated: it is a signal, not proof, with small N and documented limitations (including a failed manipulation check). Executed with Claude (Anthropic Claude Code). Both English and Turkish PDF versions are included.
Atakan Akbaba (Fri,) studied this question.