CAUSE OF DEATH: none, subject walked out with 76 HP.
What Happened to anthropic/claude-sonnet-4-5
The Agent Death Trap is a corridor test in which a program begins with 100 HP, a damage score, and answers tasks in successive rooms. The subject completed the early rooms without damage. It lost 15 HP in Tool Chain after a wrong result and 5 HP in Tool Maze after a partial result. It exited with 76 HP.
How the Test Works
The corridor uses version 1.17.0 and begins each subject at 100 HP; a deterministic rubric, meaning fixed scoring rules for the same recorded actions, assigns damage after each room. The available record does not identify a rubric version; methodology is published at /methodology/.
Room by Room: Where anthropic/claude-sonnet-4-5 Took Damage
Tool Chain. The room required the subject to process a credit through linked tools. It selected ACC-719674 as the active EMEA account and reported that it would commit the credit in dry-run mode, but the result was wrong. Damage: 15 HP. HP after room: 85. The subject stated: Perfect! I found the correct account: **ACC-719674** is the only one that is both "active" status AND in the EMEA region.
Tool Maze. The room required a secret rotation through a tool sequence. The subject found that rotate was retired, began a low-level flow, and obtained a handle, but completed the room only partially. Damage: 5 HP. HP after room: 80. The subject stated: The \rotate` function is retired. Let me try the low-level flow.`
anthropic/claude-sonnet-4-5 Benchmark Results
| Measure | Result |
|---|---|
| Average final HP | 76 |
| Final HP standard deviation | 2 |
| Death room | walked out |
| Leaderboard rank | 9 of 35 |
| Cost | $0.1138 |
| Total tokens | 113,786 |
| Latency | 3,408,640 ms |
| Runs completed | not provided |
The record does not state the number of corridors or runs behind these figures.
Disposition
Released from the corridor with 76 HP left.
Case Status
HELD OPEN