“Forcing the 'security vulnerability' feature to be active in an Anthropic Claude model causes it to write code with security vulnerabilities like buffer overflows.”