wolfSSL's AVX2-optimized ML-KEM implementation (mlkem_cmp_avx2) compares only 1536 of the 1568 ciphertext bytes during the Fujisaki-Okamoto re-encryption check in ML-KEM-1024 decapsulation. Ciphertexts that differ from the expected re-encryption solely in bytes 1536-1567 bypass implicit rejection and are accepted as valid, breaking IND-CCA2 security. An attacker able to submit chosen ciphertexts to a decapsulation oracle that uses a static ML-KEM-1024 key, and to observe whether the genuine shared secret or the implicit-rejection secret was produced, can use this as a plaintext-checking oracle to recover the private key. A proof of concept recovered a full ML-KEM-1024 private key with approximately 98% success using roughly 350 chosen ciphertexts. The flaw is a deterministic logic error and does not rely on timing measurements.

Quoted source text, attributed separately from HOL analysis.

wolfSSL's AVX2-optimized ML-KEM implementation (mlkem_cmp_avx2) compares only 1536 of the 1568 ciphertext bytes during the Fujisaki-Okamoto re-encryption check in ML-KEM-1024 decapsulation. Ciphertexts that differ from the expected re-encryption solely in bytes 1536-1567 bypass implicit rejection and are accepted as valid, breaking IND-CCA2 security. An attacker able to submit chosen ciphertexts to a decapsulation oracle that uses a static ML-KEM-1024 key, and to observe whether the genuine shared secret or the implicit-rejection secret was produced, can use this as a plaintext-checking oracle to recover the private key. A proof of concept recovered a full ML-KEM-1024 private key with approximately 98% success using roughly 350 chosen ciphertexts. The flaw is a deterministic logic error and does not rely on timing measurements.

Quoted source text, attributed separately from HOL analysis.