“A "security vulnerability" feature was discovered in Claude which, when artificially activated, causes the model to generate code containing security flaws like buffer overflows.”