CVE-2026-34159
CriticalSummary
W wersjach przed b8492 biblioteka llama.cpp, odpowiedzialna za inferencję modeli LLM, nie przeprowadzała walidacji granic w funkcji deserialize_tensor() dla pól bufora tensorów równych 0. To pozwalało nieautoryzowanym atakującym na odczyt i zapis dowolnej pamięci procesu poprzez spreparowane wiadomości GRAPH_COMPUTE.
Risk Assessment
Atakujący mógł uzyskać pełny dostęp do pamięci procesu, co prowadziło do obejścia ASLR i zdalnego wykonania kodu. Wymagany był jedynie dostęp TCP do portu serwera RPC, bez potrzeby autoryzacji.
Recommendation
Zaleca się aktualizację do wersji b8492 lub nowszej, aby usunąć tę podatność. Należy również ograniczyć dostęp do portu serwera RPC tylko do zaufanych źródeł.
Related vulnerabilities
- CVE-2026-96207Critical
Improper certificate validation in Microsoft Partner Center allows an unauthorized attacker to elevate privileges over a network.
- CVE-2026-94510Critical
Authorization bypass through user-controlled key in Microsoft Bookings allows an unauthorized attacker to elevate privileges over a network.
- CVE-2026-88131Critical
Deserialization of untrusted data in Microsoft Dataverse allows an unauthorized attacker to execute code over a network.
- CVE-2026-77900Critical
Missing authentication for a critical function in Azure App Service allows an unauthorized attacker to execute code over a network.
- CVE-2026-69435Critical
Missing authorization in Azure SRE Agent allows an authorized attacker to elevate privileges over a network.
- CVE-2026-84249Critical
IBM Guardium Data Protection 12.2 and 12.2.2 could allow a remote attacker to execute arbitrary management operations due to missing authentication for a critical function.
- CVE-2026-80381Critical
IBM Guardium Data Protection 12.0, 12.1, and 12.2 could allow a remote attacker to execute unauthorized SQL statements due to SQL injection.
- CVE-2026-75875Critical
IBM Guardium Data Protection 12.0, 12.1, and 12.2 could allow a remote attacker to execute arbitrary code due to path traversal.
- CVE-2026-107726Critical
Hazelcast prior to 5.4.5, 5.5.10, and 5.6.1 contains improper validation of data supplied by a malicious client able to connect to a cluster, allowing arbitrary reads from a cluster member's Java heap, off-heap data, and JVM process address space. The same flaw can crash cluster members and, in some Hazelcast Enterprise Edition configurations, corrupt memory with possible arbitrary code execution. This issue is fixed in versions 5.4.5, 5.5.10, 5.6.1, and 5.7.0.
- CVE-2026-107722Critical
fast-jwt from 6.2.0 until 6.3.0 can misclassify RSA public-key text as an HMAC secret when the key has non-whitespace content before its PEM header. In src/crypto.js, performDetectPublicKeyAlgorithms trims whitespace but publicKeyPemMatcher remains start-anchored, so comments, control characters, zero-width characters, or wrapper text can prevent PEM detection and reach the HMAC fallback. An attacker who knows the public key bytes can sign arbitrary HS256 claims with that public material when HS256 is inferred or allowed, resulting in authentication or authorization bypass. An asymmetric-only algorithm allowlist prevents the attack. This issue is fixed in version 6.3.0.
Original NVD description (English source)
llama.cpp is an inference of several LLM models in C/C++. Prior to version b8492, the RPC backend's deserialize_tensor() skips all bounds validation when a tensor's buffer field is 0. An unauthenticated attacker can read and write arbitrary process memory via crafted GRAPH_COMPUTE messages. Combined with pointer leaks from ALLOC_BUFFER/BUFFER_GET_BASE, this gives full ASLR bypass and remote code execution. No authentication required, just TCP access to the RPC server port. This issue has been patched in version b8492.
Vulnerability data from NVD (NIST) · CISA KEV · EPSS