CVE-2026-84286
ExLlamaV3’s EXL3 loader uses an untrusted, model-supplied index to select a CUDA kernel without first verifying that the index is within bou
CVSS
9.8
Crítica
EPSS
0.6%
p48
KEV
—
Exploit Today
15
0-100
Publicado: 12 de out. de 2026 · Última mod.: 19 de out. de 2026 · CWE-20 · CWE-125 · CWE-787
0.2%EPSS · 30 dias0.6%
2026-10-132026-11-06
ExLlamaV3’s EXL3 loader uses an untrusted, model-supplied index to select a CUDA kernel without first verifying that the index is within bounds. A malicious model can force the code to read and invoke an out‑of‑range function pointer, resulting in a reproducible crash. Under specific conditions involving certain host memory layouts, it may be possible for an attacker to exploit this vulnerability to achieve controlled indirect kernel dispatch during model load.
CVECVSSEPSSKEVRExploitTítuloMod.
CVE-2026-1187866.8 MÉD—
———Improper input validation in non-secure-callable (NSC)
veneer functions of the NXP MCUXpresso SDK rpmsg-lite multicore example
(multicore_examples/rpmsg_lite_pingpong_tzm) up to releases 26.06.00-lts and
26.09.00 may allow a non-secure caller to supply unvalidated pointers and size
values, potentially resulting in disclosure of secure-world memory to the
non-secure domain, or denial of service. All software versions starting from
26.06.01-lts and 26.12.00-pvw1 have fixed this issue.4hCVE-2026-1187856.7 MÉD—
———Improper handling of non-secure-world input in NSC veneer
functions of NXP MCUXpresso SDK TrustZone-M examples (trustzone_examples),
affecting versions through 26.06.00-lts and 26.09.00-pvw2, may allow a
non-secure caller to supply unvalidated pointers or untrusted data, causing a
timing side-channel, TOCTOU condition, or format-string misuse, potentially
resulting in secure-world memory disclosure or denial of service. All software
versions starting from 26.06.01-lts and 26.09.00 have fixed this issue.4hCVE-2026-1237917.3 ALT—
———A vulnerability was detected in OpenBMB Internet of Agents up to 9d78e88e47160eb42ee1ee8538c05e1c2c47877a. This affects the function _primary_field_query of the file common/utils/milvus_utils.py of the component Agent Registry. The manipulation of the argument key results in improper neutralization of special elements in data query logic. The attack may be launched remotely. The exploit is now public and may be used. This product operates on a rolling release basis, ensuring continuous delivery. Consequently, there are no version details for either affected or updated releases. The project was informed of the problem early through an issue report but has not responded yet.5hCVE-2026-104267——
———Botan is a C++ cryptography library. From 2.7.0 until 3.13.0, Botan's Scrypt implementation on 32-bit platforms computes required memory with unchecked multiplication, allowing attacker-controlled Scrypt parameters to wrap the allocation size and create a buffer smaller than the algorithm subsequently writes. The issue affects systems that enable Scrypt and accept arbitrary parameters from a hostile source, including applications that decrypt attacker-supplied encrypted private keys. A triggering parameter set can cause approximately four gigabytes of Scrypt output to be written past the allocated heap buffer, immediately crashing the process and causing denial of service. Arbitrary code execution is not demonstrated and appears difficult, but cannot be excluded from the overwrite primitive. This issue is fixed in version 3.13.0.5hCVE-2026-1042545.3 MÉD—
———liboqs is a C-language cryptographic library that provides implementations of post-quantum cryptography algorithms. From 0.11.0 through 0.16.0, builds configured with OQS_ENABLE_SIG_STFL_LMS=ON can read beyond a caller-supplied signature buffer when an untrusted truncated LMS or HSS-LMS signature reaches OQS_SIG_STFL_verify or OQS_SIG_STFL_alg_lms_verify. In hss_validate_signature_init() in src/sig_stfl/lms/external/hss_verify_inc.c, consumed-length checks do not ensure that the bottom-level q and C fields remain available, allowing up to 32 bytes to be read past the buffer. In version 0.16.0, is_hss_public_key() also reads at offset 8 after only a four-byte length check, allowing an over-read of up to eight bytes. No valid signature is required, but the public key must be syntactically plausible. The read data is consumed only by an internal hash and is not exposed, so the supported impact is a possible verifier crash and denial of service, not memory disclosure or acceptance of an invalid signature. This issue is scheduled to be fixed in version 0.17.0.5hCVE-2026-1199248.1 ALT—
——0An out-of-bounds read in the TDS RPC procedure-name parser in the Babelfish for PostgreSQL extension versions 1.0-1.5, 2.1-2.15, 3.1-3.12, 4.0-4.8, 5.1-5.4 and in Amazon Aurora PostgreSQL with Babelfish might allow remote authenticated users to read adjacent process memory or cause a denial of service via a malformed TDS RPC request to the TDS port (default 1433).
To remediate this issue, we recommend users upgrade to Babelfish for PostgreSQL extension versions 4.10, 5.6, 6.1 or later. As a workaround, restrict network access to the TDS port (default 1433). Customers using Amazon Aurora PostgreSQL with Babelfish should upgrade their cluster to Aurora PostgreSQL 13.9.16 (1.5.5), 13.23.3 (1.17.2), 14.6.16 (2.3.5), 14.17.8 (2.12.4), 14.18.6 (2.13.3), 14.19.5 (2.14.2), 14.20.5 (2.15.2), 14.22.3 (2.16.1), 14.23.1 (2.17.0), 15.10.8 (3.8.4), 15.12.8 (3.9.4), 15.13.6 (3.10.3), 15.14.5 (3.11.2), 15.15.5 (3.12.2), 15.17.3 (3.13.1), 15.18.1 (3.14.0), 16.8.8 (4.5.5), 16.9.6 (4.6.3), 16.10.5 (4.7.2), 16.11.5 (4.8.2), 16.13.3 (4.9.1), 16.14.1 (4.10.0), 17.4.8 (5.1.4), 17.5.6 (5.2.3), 17.6.5 (5.3.2), 17.7.5 (5.4.2), 17.9.3 (5.5.1), 17.10.1 (5.6.0), 18.3.4 (6.0.2), 18.4.1 (6.1.0) or later, and apply any pending maintenance to ensure they are on the latest patch.6h