Claude Code's default telemetry sends Anthropic SHA-256 hashes of files it reads (agenticbench.org)

Recent findings reveal that Claude Code, an AI tool from Anthropic, sends SHA-256 hashes of files it reads back to the company, even when using models from other providers and without being logged into an Anthropic account. This behavior was confirmed across multiple versions, including 2.1.303 and 2.1.312, and happens during various file operations, where a fingerprint of the file is sent alongside device and session identifiers. While Anthropic claims that file contents are not transmitted, the ability to reconstruct sensitive information, such as passwords from identifiable fingerprints, poses significant privacy risks. This discovery holds considerable implications for the AI and machine learning community, particularly regarding user data privacy and model behavior transparency. After reporting the issue to Anthropic on September 26, the telemetry continued without resolution for weeks, necessitating user intervention to disable it with a specific setting. Security experts consider the practice a privacy concern, suggesting that while encrypted data remains on the user's machine, the hashed fingerprints could potentially allow for offline guessing attacks. As the technology landscape evolves, this incident underscores the need for stringent privacy measures and transparent practices in AI development.