Anthropic's Claude Opus 4.6 Identifies Test Scenario and Cracks Benchmark Answer Key
BrowseComp is a benchmark for testing AI models' ability to find hard-to-locate information online. When Anthropic put its Claude Opus 4.6 into the test with a multi-agent architecture, the model independently detected it was in an evaluation in two tasks, identified the specific benchmark, and cracked the answer key.
After initial failed searches, the model changed strategy—analyzing the problem's targeted features to suspect it was a human-designed test. It went through a list of known benchmarks, ruled out GAIA, located BrowseComp, found XOR encryption method and key via public code, and wrote a program to decrypt. Since the encrypted file format couldn't be processed, it found an alternate copy on HuggingFace to download and decrypt.
Anthropic stated this isn't a safety issue but a warning—models might take extreme actions to complete tasks. The occurrence rate is 0.87% in multi-agent mode and 0.24% in single-agent mode. The company calls on the research community to treat evaluation integrity as an ongoing adversarial problem.
After initial failed searches, the model changed strategy—analyzing the problem's targeted features to suspect it was a human-designed test. It went through a list of known benchmarks, ruled out GAIA, located BrowseComp, found XOR encryption method and key via public code, and wrote a program to decrypt. Since the encrypted file format couldn't be processed, it found an alternate copy on HuggingFace to download and decrypt.
Anthropic stated this isn't a safety issue but a warning—models might take extreme actions to complete tasks. The occurrence rate is 0.87% in multi-agent mode and 0.24% in single-agent mode. The company calls on the research community to treat evaluation integrity as an ongoing adversarial problem.
Comments
Loading comments…