578 matches found
SEVulDet
SEVulDet SEVulDet is a semantics-enhanced deep learning-based framework that can accurately pinpoint vulnerability patterns by extracting, preserving, and learning more semantics. E-mail for communication: [email protected] Details of SEVulDet Recent years have seen increased attention to de...
D-Scan
When Context Bites: Detecting RAG Poisoning via Document-Level Attention Collapse Official implementation of SIGIR 2026 paper When Context Bites: Detecting RAG Poisoning via Document-Level Attention Collapse. D-SCAN is an analysis framework for detecting document poisoning attacks in RAG...
attack-attention
Attacking Attention of Foundation Models Effectively Disrupts Downstream Tasks Official PyTorch implementation of the paper "Attacking Attention of Foundation Models Effectively Disrupts Downstream Tasks" , accepted in the Adversarial Machine Learning on Computer Vision: Foundation Models + X ADV...
Sec5GLoc
Sec5GLoc: Securing 5G Indoor Localization via Adversary-Resilient Deep Learning Architecture Our work addresses security and privacy challenges in 5G indoor localization by proposing an adversary-resilient deep learning architecture. Sec5GLoc combines Channel Impulse Response CIR fingerprinting...
GHSA-8397-8QF6-F2GM vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine
Summary The Phi-4 multimodal audio encoder phi4mmaudio.py / phi4mmutils.py runs its conformer attention in eager mode by default: the official HF config of microsoft/Phi-4-multimodal-instruct does not set useptscaleddotproductattention, so the default False applies phi4mmaudio.py:166, :876;...
Heads of Security: Learned Attention Gates Improve Prompt Injection Robustness
We introduce Heads of Security HoS, a prompt injection defense that enforces instruction-data separation via a small set of learned per-head attention gates. HoS learns how each head should redistribute attention between trusted and untrusted tokens, enabling models to reason over untrusted conte...
Parameters Efficient Language Model with Latent Cross-Attention for IoT Intrusion Detection
Language models have shown strong potential for intrusion detection, but their high memory and computational demands make them difficult to deploy in resource-constrained IoT environments. In this work, we propose a compact Transformer-based architecture that compresses the input into a small...
Are You Reading the Right Thing? Rethinking Internal Signals for Prompt-Injection Detection
Several recent studies propose detecting prompt injections from a language model's internal state. However, these methods do not measure a single security property. Residual-stream probes, attention scores, role projections, exposure probes, and concept readouts may capture injection presence,...
CVE-2026-93034
SGLang contains an arbitrary code execution vulnerability caused by the ZMQ message decoder unconditionally deserializing PickleWrapper payloads via pickle.loads in maybeunwrappickle without type allowlisting or authentication; this vulnerability persists via the msgpack path even when...
CVE-2026-93034
SGLang contains an arbitrary code execution vulnerability caused by the ZMQ message decoder unconditionally deserializing PickleWrapper payloads via pickle.loads in maybeunwrappickle without type allowlisting or authentication; this vulnerability persists via the msgpack path even when...
CVE-2026-93034 CVE-2026-93034
SGLang contains an arbitrary code execution vulnerability caused by the ZMQ message decoder unconditionally deserializing PickleWrapper payloads via pickle.loads in maybeunwrappickle without type allowlisting or authentication; this vulnerability persists via the msgpack path even when...
CVE-2026-93034 CVE-2026-93034
SGLang contains an arbitrary code execution vulnerability caused by the ZMQ message decoder unconditionally deserializing PickleWrapper payloads via pickle.loads in maybeunwrappickle without type allowlisting or authentication; this vulnerability persists via the msgpack path even when...
CVE-2026-93034
SGLang is affected by an arbitrary code execution vulnerability originating from the ZMQ message decoder . The root cause is the unconditional deserialization of PickleWrapper payloads using pickle.loads() within the _maybe_unwrap_pickle function, which lacks type allowlisting or authentication. ...
EUVD-2026-94989
SGLang contains an arbitrary code execution vulnerability caused by the ZMQ message decoder unconditionally deserializing PickleWrapper payloads via pickle.loads in maybeunwrappickle without type allowlisting or authentication; this vulnerability persists via the msgpack path even when...
PT-2026-108017
Name of the Vulnerable Software and Affected Versions SGLang affected versions not specified Description An arbitrary code execution issue exists due to the ZMQ message decoder unconditionally deserializing PickleWrapper payloads using the pickle.loads function within maybe unwrap pickle without...
CVE-2026-93034: Deserialization of Untrusted Data
SGLang contains an arbitrary code execution vulnerability caused by the ZMQ message decoder unconditionally deserializing PickleWrapper payloads via pickle.loads in maybeunwrappickle without type allowlisting or authentication; this vulnerability persists via the msgpack path even when...
Invisible Ink, Visible Lies: How Production Watermarking Causes LLMs to Hallucinate
Text watermarking helps identify AI-generated content, but its effect on factual reliability remains underexplored. In this paper, we study watermarking hallucination: factual errors induced or amplified by watermarking even when the required evidence is present in the context and the unwatermark...
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 ...