The US Government Has Asked Anthropic to Make Its LLMs Unhackable

The government asked Anthropic to make large language models unhackable. The government did not define unhackable. Anthropic has not provided a definition. Both parties have proceeded on the assumption that the conversation will clarify what was meant, or that clarification is not necessary, or that both states are equivalent.

This is how security requirements are written when the requirement itself is not fully specified by anyone involved. The term unhackable means different things at different times to different security frameworks. The government is now asking Anthropic what the government was asking. This is not unusual.

Anthropus will produce something. The government will review something. An inspector general's office will eventually ask both sides what the original question was. The answer will be filed.