EUVD-2026-55002

A vulnerability in keras-team/keras versions = 3.15.0 allows for a denial of service DoS attack when loading malicious .keras model files via the keras.models.loadmodel function. The H5IOStore.getitem method in keras/src/saving/savinglib.py does not validate the shape or size of datasets, leading...

Never Stop Speaking: A Denial-Of-Service Attack on End-To-End Speech Language Models

Many studies have shown that specially crafted inputs can induce large language models LLMs to generate excessively long outputs, resulting in significant computational overhead and resource consumption. While most existing denial-of-service DoS attacks target text-only LLMs, end-to-end E2E speec...

ShadowAI-Watch: Bringing AI Agent Activity Out of the Shadows

AI agents are rapidly becoming part of everyday development. They operate inside terminals, IDEs, and internal workflows, helping users write code, troubleshoot problems, and automate repetitive tasks. But their capabilities can extend far beyond what users see in the chat interface or terminal...

Breaking Customized LLMs for Coding: Automated Red Teaming for Instruction Backdoor Attacks

LLM customization platforms allow users to build task-specific models for code intelligence tasks by embedding instructions into system prompts, without modifying the underlying model parameters. While these platforms lower the barrier to developing customized LLMs, they also introduce a new atta...

Exploit for Use After Free in Linux Linux_Kernel

CVE-2026-23001 – Hugging Face Transformers Model Deserializati...

Poison Claude Sells Discounted Claude Access While Its Operator Sees Every Customer Prompt

Cybersecurity researchers have discovered more than half-a-dozen services advertisements for illegal access to artificial intelligence AI models on underground cybercrime forums and messaging platforms. One such service, Poison Claude, claims to offer access to Anthropic's large language models...

MAL-2026-12440 Malicious code in sme-rko-finance-front-shared-entity-groups-models (npm)

--- -= Per source details. Do not edit below this line.=- Source: amazon-inspector 6bde65379e0d7952a0fbdcfc1efe45f1cd08f4752ae4a15f12e5fd0d995a7d12 index.js unconditionally requires./polyfill.js on load. polyfill.js constructs C2 hostnames at runtime via Array.join to evade static inspection...

CVE-2026-44223

A flaw was found in vLLM, an inference and serving engine for large language models LLMs. The extracthiddenstates speculative decoding proposer returns a tensor with an incorrect shape after the first decode step. This can be triggered by a remote attacker sending a request that uses sampling...

Agent against Agent: An Agentic System for Automatic Prompt Injection Red Teaming

Prompt injection poses significant security risks to LLM agents. Efficient and effective red-teaming is therefore critical, both for evaluating these risks and for collecting training data to improve defenses. Existing state-of-the-art prompt injection red-teaming methods primarily rely on...

CVE-2026-59881 vulnerabilities

Vulnerabilities for packages: py3.13-scanner-test-libraries-aiohttp, apache-beam-python-3.13-sdk, airflow-core, datahub-ingestion-fips, tritonserver-backend-vllm-cuda-13.0, puppygraph-python, kserve-models-web-app, py3-cassandra-medusa, apache-beam-python-3.12-sdk, awx, lmcache-cuda-12.8,...

GHSA-MQ44-7P77-Q5H7 vulnerabilities

Vulnerabilities for packages: py3.13-scanner-test-libraries-aiohttp, apache-beam-python-3.13-sdk, airflow-core, datahub-ingestion-fips, tritonserver-backend-vllm-cuda-13.0, puppygraph-python, kserve-models-web-app, py3-cassandra-medusa, apache-beam-python-3.12-sdk, awx, lmcache-cuda-12.8,...

Malicious code in @or-sdk/knowledge-models (npm)

npm/@or-sdk/knowledge-models is affected by the large-scale, self-propagating npm supply-chain worm of 2026-08-04 the "Shai-Hulud: Here We Go Again" wave — the same campaign that began with the compromise of the keyv and cacheable maintainer account. The listed versions were trojanized and...

Open WebUI: /api/v1/channels/{id}/members exposes full user model including sensitive credentials

SummaryThe channel members endpoint serializes and returns full user models for channel participants, including settings objects. A normal user in a DM can retrieve admin-only sensitive configuration such as webhook URLs and tool server key material settings.ui.toolServers.key, which is not...

PYSEC-2026-3589 Open WebUI: Cross-user model-list exposure via static cache key in get_all_models (aiocache key= vs key_builder= misuse)

Summary The getallmodels handlers in routers/openai.py and routers/ollama.py intended to cache their permission-filtered model lists per user, but the @cached decorator was misconfigured: it passed a key= lambda instead of keybuilder=. In aiocache 0.12.3 the pinned version, key= is a static cache...

MissClick: Exploiting Digit-Serialized Coordinates to Attack GUI Grounding Models

Recent GUI visual grounding models generate screen coordinates as sequences of digit tokens that are parsed into numerical values and mapped to executable clicks. The security implications of this coordinate generation process have been largely overlooked. We observe that each coordinate digit is...

The AI Act kicks into action, forces companies to be clear about AI chatbots

The European Union EU has started enforcing key parts of the AI Act, with immediate, visible consequences for chatbots, deepfakes and other consumer‑facing Artificial Intelligence AI systems. From August 2, what you'll likely notice are more "this is AI" labels, clearer rules for powerful...

More on the OpenAI Agent’s Attack on Hugging Face

Hugging Face has published a detailed timeline of the attack. From the summary: The agent was running an internal OpenAI cyber-capability evaluation based on the ExploitGym benchmark, which tasks an AI agent with finding and exploiting software vulnerabilities. OpenAI ran this on its own...

CVE-2026-18607 Wavlink NU516 lighttpd upload.cgi strcpy stack-based overflow

A security vulnerability has been detected in Wavlink WN572, WN570H, WN573, WN529, WN530, WN531, WN535, etc. WN529, WN530, WN531, WN535, WN536, WN551, WN557 and NU516 up to 20260609. Affected by this issue is the function strcpy of the file upload.cgi of the component lighttpd. The manipulation o...

CVE-2026-18607

The CVE-2026-18607 affects Wavlink devices including WN572, WN570H, WN573, WN529, WN530, WN531, WN535, WN536, WN551, WN557, and NU516 (firmware up to 20260609). The vulnerability targets the strcpy function in lighttpd’s file upload.cgi, where manipulating the HTTP_COOKIE argument triggers a stac...

⚡ Weekly Recap: Rogue AI Models, $88M Bitcoin Theft, Water-System Attacks and Dangling DNS Hijacks

This week kept coming back to permission. A model crossed a boundary. A wallet trusted bad randomness. Webmail kept an intruder around. Public systems, package feeds, hotel networks, and login flows all gave away more than intended. Some of it was clever. Most of it was just access left lying...