Latent Fusion Jailbreak: Blending Harmful and Harmless Representations to Elicit Unsafe LLM Outputs

Large language models LLMs demonstrate impressive capabilities in various language tasks but are susceptible to jailbreak attacks that circumvent their safety alignments. This paper introduces Latent Fusion Jailbreak LFJ, a representation-based attack that interpolates hidden states from harmful...

AuthPrint: Fingerprinting Generative Models against Malicious Model Providers

Generative models are increasingly adopted in high-stakes domains, yet current deployments offer no mechanisms to verify the origin of model outputs. We address this gap by extending model fingerprinting techniques beyond the traditional collaborative setting to one where the model provider may a...

CVE-2025-32874

An issue was discovered in Kaseya Rapid Fire Tools Network Detective through 2.0.16.0. A vulnerability exists in the EncryptionUtil class because symmetric encryption is implemented in a deterministic and non-randomized fashion. The method Encryptbyte clearData derives both the encryption key and...

CVE-2025-32874

CVE-2025-32874 affects Kaseya Rapid Fire Tools Network Detective up to version 2.0.16.0. The issue is in the EncryptionUtil class where symmetric encryption is implemented deterministically; the key and IV are derived from a fixed, hardcoded input using a static salt. As a result, identical plain...

Expected Behavior Violation

Overview llama-index is an Interface between LLMs and your data Affected versions of this package are vulnerable to Expected Behavior Violation via the DocugamiReader class. An attacker can cause loss of important document content, disrupt parent-child chunk hierarchies, and lead to inaccurate AI...

CVE-2025-6211 MD5 Hash Collision in run-llama/llama_index

A vulnerability in the DocugamiReader class of the run-llama/llamaindex repository, up to version 0.12.28, involves the use of MD5 hashing to generate IDs for document chunks. This approach leads to hash collisions when structurally distinct chunks contain identical text, resulting in one chunk...

PT-2025-31076

Name of the Vulnerable Software and Affected Versions Linux kernel affected versions not specified Description The insn rw emulate bits function within the Comedi subsystem does not properly handle cases where insn-n is 0 for INSN READ and INSN WRITE instructions. This can lead to the function...

SecureT2I: No More Unauthorized Manipulation on AI Generated Images from Prompts

Text-guided image manipulation with diffusion models enables flexible and precise editing based on prompts, but raises ethical and copyright concerns due to potential unauthorized modifications. To address this, we propose SecureT2I, a secure framework designed to prevent unauthorized editing in...

CVE-2025-48466

Successful exploitation of the vulnerability could allow an unauthenticated, remote attacker to send Modbus TCP packets to manipulate Digital Outputs, potentially allowing remote control of relay channel which may lead to operational or safety risks...

CVE-2025-48466

Successful exploitation of the vulnerability could allow an unauthenticated, remote attacker to send Modbus TCP packets to manipulate Digital Outputs, potentially allowing remote control of relay channel which may lead to operational or safety risks...

CVE-2025-48466

Successful exploitation of the vulnerability could allow an unauthenticated, remote attacker to send Modbus TCP packets to manipulate Digital Outputs, potentially allowing remote control of relay channel which may lead to operational or safety risks...

CVE-2025-48466 Modbus Command Injection without Authentication

Successful exploitation of the vulnerability could allow an unauthenticated, remote attacker to send Modbus TCP packets to manipulate Digital Outputs, potentially allowing remote control of relay channel which may lead to operational or safety risks...

UCD: Unlearning in LLMs Via Contrastive Decoding

Machine unlearning aims to remove specific information, e.g. sensitive or undesirable content, from large language models LLMs while preserving overall performance. We propose an inference-time unlearning algorithm that uses contrastive decoding, leveraging two auxiliary smaller models, one train...

SoK: the Privacy Paradox of Large Language Models: Advancements, Privacy Risks, and Mitigation

Large language models LLMs are sophisticated artificial intelligence systems that enable machines to generate human-like text with remarkable precision. While LLMs offer significant technological progress, their development using vast amounts of user data scraped from the web and collected from...

HauntAttack: When Attack Follows Reasoning As a Shadow

Emerging Large Reasoning Models LRMs consistently excel in mathematical and reasoning tasks, showcasing exceptional capabilities. However, the enhancement of reasoning abilities and the exposure of their internal reasoning processes introduce new safety vulnerabilities. One intriguing concern is:...

CVE-2024-2445

Mattermost Jira plugin versions shipped with Mattermost versions 8.1.x before 8.1.10, 9.2.x before 9.2.6, 9.3.x before 9.3.2, and 9.4.x before 9.4.3 fail to escape user-controlled outputs when generating HTML pages, which allows an attacker to perform reflected cross-site scripting attacks agains...

CVE-2024-7816

The Gixaw Chat WordPress plugin through 1.0 does not have CSRF check in some places, and is missing sanitisation as well as escaping, which could allow attackers to make logged in admin add Stored XSS payloads via a CSRF attack...

CVE-2021-39947

In specific circumstances, trace file buffers in GitLab Runner versions up to 14.3.4, 14.4 to 14.4.2, and 14.5 to 14.5.2 would re-use the file descriptor 0 for multiple traces and mix the output of several jobs...

CVE-2024-58135

Mojolicious versions from 7.28 through 9.45 for Perl will generate weak HMAC session cookie secrets via "mojo generate app" by default. When creating a default app skeleton with the "mojo generate app" tool, a weak secret is written to the application's configuration file using the insecure rand...

Protocol-Agnostic and Data-Free Backdoor Attacks on Pre-Trained Models in RF Fingerprinting

While supervised deep neural networks DNNs have proven effective for device authentication via radio frequency RF fingerprinting, they are hindered by domain shift issues and the scarcity of labeled data. The success of large language models has led to increased interest in unsupervised pre-train...