An issue was discovered in Mbed TLS before 2.28.1 and 3.x before 3.2.0. In some configurations, an unauthenticated attacker can send an invalid ClientHello message to a DTLS server that causes a heap-based buffer over-read of up to 255 bytes. This can cause a server crash or possibly information disclosure based on error responses. Affected configurations have MBEDTLS_SSL_DTLS_CLIENT_PORT_REUSE enabled and MBEDTLS_SSL_IN_CONTENT_LEN less than a threshold that depends on the configuration: 258 bytes if using mbedtls_ssl_cookie_check, and possibly up to 571 bytes with a custom cookie check function.
In Mbed TLS before 3.1.0, psa_aead_generate_nonce allows policy bypass or oracle-based decryption when the output buffer is at memory locations accessible to an untrusted application.
An issue was discovered in Mbed TLS before 2.24.0. The verification of X.509 certificates when matching the expected common name (the cn argument of mbedtls_x509_crt_verify) with the actual certificate name is mishandled: when the subjecAltName extension is present, the expected name is compared to any name in that extension regardless of its type. This means that an attacker could impersonate a 4-byte or 16-byte domain by getting a certificate for the corresponding IPv4 or IPv6 address (this would require the attacker to control that IP address, though).
In Trusted Firmware Mbed TLS 2.24.0, a side-channel vulnerability in base64 PEM file decoding allows system-level (administrator) attackers to obtain information about secret RSA keys via a controlled-channel and side-channel attack on software running in isolated environments that can be single stepped, especially Intel SGX.