attack-attention

파운데이션 모델의 어텐션을 공격하면 다운스트림 태스크를 효과적으로 붕괴시킨다 논문 "Attacking Attention of Foundation Models Effectively Disrupts Downstream Tasks" 의 공식 PyTorch 구현체로, CVPR 2025 의 Adversarial Machine Learning on Computer Vision: Foundation Models + X ADVML 워크숍에 채택되었습니다. "Attacking Attention of Foundation Models...

D-Scan

コンテキストが牙をむくとき：文書レベルの注意崩壊によるRAGポイズニングの検出 D-SCANは、RAG（Retrieval-Augmented Generation）システムにおける文書ポイズニング攻撃を検出するための分析フレームワークです。生成中にLLMの内部状態（トークン確率と注意重み）を収集し、多次元特徴量を抽出し、クリーンな検索文書とポイズニングされた検索文書を区別する分類器を訓練します。 1. インストールと要件 前提条件 Python = 3.9 CUDA対応GPU（8Bパラメータモデルの推論には24GB以上のVRAMを推奨） モデルの準備 このプロジェクトはデフォルトで...

SEVulDet

SEVulDet SEVulDet는 더 많은 시맨틱을 추출, 보존 및 학습하여 취약점 패턴을 정확하게 찾아내는 시맨틱 강화 딥러닝 기반 프레임워크입니다. 이메일 문의: [email protected] SEVulDet 상세 정보 최근 몇 년간 신경망을 활용하여 취약점 패턴을 식별하는 딥러닝 기반 취약점 탐지 프레임워크에 대한 관심이 증가했습니다. 많은 노력이 있었지만, 기존 접근 방식은 실제로 정확도가 낮습니다. 이전 연구들은 소스 코드에서 시맨틱을 포괄적으로 캡처하지 못하거나 신경망의 적절한 설계를 채택하지 못합니다. ...

Sec5GLoc

Sec5GLoc: 적대적 공격에 강인한 딥러닝 아키텍처를 통한 5G 실내 측위 보안 본 연구는 적대적 공격에 강인한 딥러닝 아키텍처를 제안하여 5G 실내 측위의 보안 및 프라이버시 문제를 해결합니다. Sec5GLoc은 채널 임펄스 응답CIR 핑거프린팅과 도착 시간 차TDoA 피처 및 알려진 앵커 위치를 결합하며, 하이브리드 CNN과 멀티-헤드 어텐션 네트워크가 이를 처리합니다. 이 설계는 위치 스푸핑 및 적대적 신호 변조와 같은 위협에 대한 견고성을 향상시킵니다. 그림 1. 적대적 신호 교란 및 물리 계층 위협 하에서 안전하고...

CVE-2026-91073 Subscribe Forms 1.4.1 - 1.6.2 - Author+ Stored XSS via Attention Effect Form Setting

The Subscribe Forms WordPress plugin before 1.6.3 does not sanitise and escape one of its form settings before outputting it in a page, allowing authenticated users with the Author role and above to perform Stored Cross-Site Scripting attacks that execute in the browser of any visitor who views a...

CVE-2026-91073 Subscribe Forms 1.4.1 - 1.6.2 - Author+ Stored XSS via Attention Effect Form Setting

The Subscribe Forms WordPress plugin before 1.6.3 does not sanitise and escape one of its form settings before outputting it in a page, allowing authenticated users with the Author role and above to perform Stored Cross-Site Scripting attacks that execute in the browser of any visitor who views a...

CVE-2026-91073

The Subscribe Forms WordPress plugin (versions 1.4.1-1.6.2 ) is vulnerable to a Stored Cross-Site Scripting (XSS) attack. The root cause is the failure to properly sanitise and escape specific form settings before they are outputted on a page. This allows authenticated users with the Author role ...

Astra Linux – Vulnerability found in Linux 5.10, Linux 6.1, Linux, Linux 5.15

In the Linux kernel, the following vulnerability has been resolved: usb: typec: bus: verify partner exists in typecaltmodeattention Some USB hubs will negotiate DisplayPort Alt mode with the device. However, they will then negotiate a data role swap after entering the Alt mode. This data role swa...

EUVD-2026-48988

In the Linux kernel, the following vulnerability has been resolved: Input: synaptics-rmi4 - bound the F30 keymap to the GPIO/LED count rmif30mapgpios allocates gpioledkeymap with mingpioledcount, TRACKSTICKRANGEEND == at most 6 entries, but rmif30attention iterates the full f30-gpioledcount devic...

Symlink Attack

Overview flash-attention is a Flash Attention2 operator on Huawei Ascend 910A. Affected versions of this package are vulnerable to Symlink Attack via the downloadandcopy function in hopper/setup.py when extracting archives without validating symlinks or filtering tar members. An attacker can...

flash-attention contains an insecure deserialization vulnerability in its checkpoint loading mechanism

The flash-attention training framework thru commit e724e2588cbe754beb97cf7c011b5e7e34119e62 2025-13-04 contains an insecure deserialization vulnerability CWE-502 in its checkpoint loading mechanism. The loadcheckpoint function in checkpoint.py and the checkpoint loading code in eval.py use...

PYSEC-2026-2477 flash-attention contains an insecure deserialization vulnerability in its checkpoint loading mechanism

The flash-attention training framework thru commit e724e2588cbe754beb97cf7c011b5e7e34119e62 2025-13-04 contains an insecure deserialization vulnerability CWE-502 in its checkpoint loading mechanism. The loadcheckpoint function in checkpoint.py and the checkpoint loading code in eval.py use...

CVE-2026-46177

In the Linux kernel, the following vulnerability has been resolved: ipmi: Add limits to event and receive message requests The driver would just fetch events and receive messages until the BMC said it was done. To avoid issues with BMCs that never say they are done, add a limit of 10 fetches at a...

UBUNTU-CVE-2026-46177

In the Linux kernel, the following vulnerability has been resolved: ipmi: Add limits to event and receive message requests The driver would just fetch events and receive messages until the BMC said it was done. To avoid issues with BMCs that never say they are done, add a limit of 10 fetches at a...

CVE-2026-46177 ipmi: Add limits to event and receive message requests

In the Linux kernel, the following vulnerability has been resolved: ipmi: Add limits to event and receive message requests The driver would just fetch events and receive messages until the BMC said it was done. To avoid issues with BMCs that never say they are done, add a limit of 10 fetches at a...

CVE-2026-46177

The CVE-2026-46177 issue affects the Linux kernel IPMI driver. It describes a vulnerability where the driver could continuously fetch events and receive messages from the BMC (or become stuck) due to the BMC not signaling completion or the attn bit getting stuck. The documented fix limits event/m...

EUVD-2026-32804

In the Linux kernel, the following vulnerability has been resolved: ipmi: Add limits to event and receive message requests The driver would just fetch events and receive messages until the BMC said it was done. To avoid issues with BMCs that never say they are done, add a limit of 10 fetches at a...

PT-2026-44300

Name of the Vulnerable Software and Affected Versions Linux kernel affected versions not specified openSUSE Tumbleweed versions prior to kernel-devel-7.0.11-1.1 Description The ipmi driver fails to implement limits on event and receive message requests, continuing to fetch data until the Baseboar...

Dissecting the Black Box: Circuit-Level Analysis of LLM Vulnerability Detection

Large language models LLMs can detect software vulnerabilities, but how do they actually identify vulnerable code? We address this question using mechanistic interpretability; analyzing the internal computations of a neural network to understand its reasoning process.Using Circuit Tracer on...

CVE-2026-44075

A missing break statement in DSI OpenSession processing in Netatalk 1.5.0 through 4.4.2 causes a DSIOPTATTNQUANT switch case to fall through into DSIOPTSERVQUANT, resulting in unintended session option handling that may allow a remote attacker to cause a minor service disruption via crafted DSI...