Evaluating Frontier AI Agents As Autonomous Clinical Security Auditors

Clinical AI models can expose patients to harm when adversarial vulnerabilities go undetected, yet formal security auditing requires statistical expertise, specialized tools, and significant time. We present an open evaluation task, built on METR Task Standard v0.3.0, that tests whether frontier ...

Open WebUI: Forged model meta.knowledge allows cross-user file read and deletion

SummaryOpen WebUI lets a user who can create, update, or import workspace models store arbitrary meta.knowledge entries on their model without checking whether they own or can read the referenced files. Open WebUI then treats meta.knowledge entries of type file as an authorization source in two...

PYSEC-2026-2761 Open WebUI: Forged model meta.knowledge allows cross-user file read and deletion

Summary Open WebUI lets a user who can create, update, or import workspace models store arbitrary meta.knowledge entries on their model without checking whether they own or can read the referenced files. Open WebUI then treats meta.knowledge entries of type file as an authorization source in two...

PYSEC-2026-2757 Open WebUI: Stored XSS to Account Takeover via Model Profile Images

Stored XSS to Account Takeover via Model Profile Images in Open WebUI Affected: Open WebUI tags. On the output side, users.py added a MIME allowlist check and X-Content-Type-Options: nosniff. The fix was applied to UserUpdateForm, UpdateProfileForm, and later to ChannelWebhookForm. Three models...

PYSEC-2026-2741 Open WebUI Exposes System Prompt to Regular User [Non-Admin]

Summary A regular user non-admin can view the system prompt of the model which is set by an admin. Details When a regular user non-admin logs into the application, a http://IP:8080/api/models? web request is initiated by the application and in response, it reveals the system prompt of available...

Open WebUI's Model Import Overwrites Any Model Without Ownership Check

Model Import Overwrites Any Model Without Ownership Check Affected ComponentModel import endpoint:- backend/openwebui/routers/models.py lines 254-308, importmodels Affected VersionsCurrent main branch commit 6fdd19bf1 and likely all versions with model import functionality. DescriptionThe POST...

PYSEC-2026-2749 Open WebUI's Model Import Overwrites Any Model Without Ownership Check

Model Import Overwrites Any Model Without Ownership Check Affected Component Model import endpoint: - backend/openwebui/routers/models.py lines 254-308, importmodels Affected Versions Current main branch commit 6fdd19bf1 and likely all versions with model import functionality. Description The POS...

PYSEC-2026-2734 Open WebUI's responses passthrough endpoint lacks access control authorization

Summary The /responses endpoint in the OpenAI router accepts any authenticated user and forwards requests directly to upstream LLM providers without enforcing per-model access control. While the primary chat completion endpoint generatechatcompletion checks model ownership, group membership, and...

PYSEC-2026-2719 Open WebUI's Base Model Routing Bypasses Access Control via Model Chaining

Base Model Routing Bypasses Access Control via Model Chaining Affected Component Model chaining via basemodelid: - backend/openwebui/routers/models.py lines 170-214, createnewmodel - backend/openwebui/routers/models.py lines 254-308, importmodels - backend/openwebui/main.py lines 1696-1711, base...

InstructLab Includes Functionality from Untrusted Control Sphere

A flaw was found in InstructLab. The linuxtrain.py script hardcodes trustremotecode=True when loading models from HuggingFace. This allows a remote attacker to achieve arbitrary Python code execution by convincing a user to run ilab train/download/generate with a specially crafted malicious model...

PYSEC-2026-2522 InstructLab Includes Functionality from Untrusted Control Sphere

A flaw was found in InstructLab. The linuxtrain.py script hardcodes trustremotecode=True when loading models from HuggingFace. This allows a remote attacker to achieve arbitrary Python code execution by convincing a user to run ilab train/download/generate with a specially crafted malicious model...

Securing LLMs in the Wild: Privacy and Security Challenges at the Edge

Large Language Models LLMs are rapidly moving from research settings into the wild, deployed on enterprise infrastructure, personal devices, and edge platforms. While cloud deployments offer scalable compute, concerns over data sovereignty, compliance, latency, and third-party dependence are...

The vulnerability of the Open WebUI web interface, related to the lack of permission checking, allows a violator to gain read access, modify data, and increase their privileges.

The vulnerability of the Open WebUI web interface relates to the lack of permission checking for workspace.models.import. Exploiting this vulnerability can allow an attacker, operating remotely, to gain read access, modify data, and enhance their privileges by sending a specially crafted POST...

PT-2026-59399

Summary An internal-only bypass filter parameter is exposed on the /openai/chat/completions and /ollama/api/chat HTTP endpoints via FastAPI query string binding, allowing any authenticated user to append ?bypass filter=true and bypass model access control checks to invoke admin-restricted models...

PT-2026-59370

Summary When setting model permissions so that a group has read access to it, intending for other users to use it, those users also can read the model's system prompt. However users may consider their system prompt confidential, so we consider this a security issue. Compare...

PT-2026-59205

A vulnerability in the TFSMLayer class of the keras package, version 3.13.0, allows attacker-controlled TensorFlow SavedModels to be loaded during deserialization of .keras models, even when safe mode=True. This bypasses the security guarantees of safe mode and enables arbitrary attacker-controll...

AMT-X: Phase-Structured Multi-Turn Red-Teaming with Checklist-Gated Evaluation

Safety evaluation of large language models LLMs relies largely on single-turn attack datasets and single-judge scoring, underestimating risk from adaptive multi-turn adversaries and reporting a single success rate that does not separate partially actionable outputs from those carrying complete...

Agent Hacks Agent: Autoresearch for Production-Agent Red-Teaming

Production LLM agents such as Claude Code and Codex operate over untrusted content, files, commands, and workspace state, making safety failures directly actionable. Red-teaming must therefore keep pace with evolving models and tools. Existing approaches mainly optimize attack success and preserv...

PT-2026-59383

Summary A regular user non-admin can view the system prompt of the model which is set by an admin. Details When a regular user non-admin logs into the application, a http://IP:8080/api/models? web request is initiated by the application and in response, it reveals the system prompt of available...

Large Language Models in Misinformation Ecosystems: Misuse, Defense, and Vulnerability

Large language models LLMs have transformed misinformation from a primarily content-centric problem into a broader ecosystem-level security challenge. When misused, LLMs create risks beyond false content generation, enabling attacks on the social contexts, evidence sources, retrieval corpora, and...