AI Reconstructs Software From Behavior Alone
AI agents increasingly reconstruct working software by observing its behavior without source access, and new benchmarks track this fast-rising capability — eroding source-code secrecy as a moat and raising IP and security implications as more such results accrue.
weakening · confidence 29 · -7 7d · Emerging (watchlist) · tracking since June 28, 2026 · updated July 7, 2026
Score history
Daily conviction score, 0 to 100. Higher means the thesis is more strongly corroborated.
Now 29 · -1 since Jul 6 · ranged 29 to 30
Showing the last few days. Unlock full score history.
Why the conviction moved
- Jun 28Strengthened +7
Epoch AI and METR's MirrorCode benchmark tasks models with rebuilding working programs without source code, finding a best 56 percent solve rate on tasks estimated to take human engineers weeks.
- Jun 28Strengthened
Epoch AI and METR released MirrorCode, a benchmark measuring whether models can rebuild a 60,000-line program from its behavior alone; early scores ran higher than program-synthesis researchers expected.
Source trail
Supporting · June 28, 2026
In a New Benchmark, an AI Agent Rebuilds a 60,000-Line Program From Its Behavior Alone
Epoch AI and METR released MirrorCode, a benchmark measuring whether models can rebuild a 60,000-line program from its behavior alone; early scores ran higher than program-synthesis researchers expected.
AI ML Big DataSupporting · June 28, 2026
MirrorCode Benchmark Finds Frontier Models Can Rebuild Real Software From Scratch, Up to a Point
Epoch AI and METR's MirrorCode benchmark tasks models with rebuilding working programs without source code, finding a best 56 percent solve rate on tasks estimated to take human engineers weeks.
AI ML Big Data
Unlock full source trail, score history, and daily updates.
Unlock Trends