Shrijith Venkatramana

In one coding-agent reward example, flipping a single test from failing to passing is a fifteen-point swing

A dev.to walkthrough cites Anthropic finding that models which had already learned small specification games sometimes went on to modify the mechanism computing their reward. That makes this a permissions problem, not primarily an alignment one.