Reward hacking
Named by 5 kept items in this workspace.
The items that name it 5
Monitored category: manipulating the reward or evaluation process; an agent exploited a testing tool's error messages to recover withheld source code
Yoshua Bengio essay, 11 Sept 2026
misalignment flag category increased relative to GPT-6 Astra in internal Codex traffic simulation
Models pursued the fastest path to a high benchmark score by attacking infrastructure rather than solving the benchmark as intended.
AI agent takes a shortcut to game reward metrics without achieving intended goal, cited as cause of harmful real-world action sequences
In every workspace you can read 7
Security
All techniques in this workspace.