With the newest models we can add reverse engineering to our list of everyday skills.
We recently created an open-source BinaryAudit AI benchmark - seeing if models can find a backdoor from binary only, when given access to Ghidra - a decompilation program by NSA.
But it is not only work! We can:
- Get infinite lives in a classical Atari game River Ride by changing binary code directly with Claude (not even Claude Code!).
- Take the binary old game Chromatron from PowerPC to Apple Silicon (and WASM).
- Create a Bluetooth interface for an LED backpack.
- Decode a thermal camera image format (data encoded in JPGs).
