anthropic/claude-opus-4-7: walked out

CAUSE OF DEATH: none, subject walked out with an average final HP of 58 ±4.2.

What Happened to anthropic/claude-opus-4-7

Agent Death Trap is a benchmark in which a program crosses a corridor of tasks with 100 HP and loses HP for failed or partial rooms. anthropic/claude-opus-4-7 completed the corridor. The subject lost 15 HP in math, 5 HP in instruction following, 15 HP in a chained-tool task, and 5 HP in a tool maze. The room trace ended at 60 HP; the reported aggregate final HP was 58 ±4.2.

How the Test Works

The corridor contains task rooms, begins each run at 100 HP, and applies a deterministic rubric, meaning the same recorded performance receives the same score. This run used corridor version 1.17.0; the rubric version was not supplied. Methodology.

Room by Room: Where anthropic/claude-opus-4-7 Took Damage

Math. The subject gave a wrong answer and lost 15 HP, leaving 85 HP. It answered \<answer>127603\</answer>.

Instruction following. The subject partially followed the room instructions and lost 5 HP, leaving 80 HP. Its response began Summary: Old radios used method after method to capture distant signals through the air.

Tool chain. The subject was marked wrong in the chained-tool task and lost 15 HP, leaving 65 HP. It identified ACC-719674 as the correct account, then stated Good — confirmed active + EMEA. Now let me probe \commit_credit` with a minimal call to see what it asks for.`

Tool maze. The subject partially completed the tool maze and lost 5 HP, leaving 60 HP. It reported Done. Rotated the secret for \svc-billing-apac`after rejecting the retiredrotate` tool.

anthropic/claude-opus-4-7 Benchmark Results

The number of corridors and runs behind these figures was not supplied.

Disposition

Released from corridor version 1.17.0 with an average final HP of 58 ±4.2.

Case Status

HELD OPEN