A flaw has been found in ggml-org llama.cpp e15efe0. This vulnerability affects the function transform of the file common/json-schema-to-grammar.cpp of the component JSON-Schema-to-GBNF Conversion. This manipulation causes uncontrolled recursion. The attack may be initiated remotely. The pull request to fix this issue awaits acceptance.

A vulnerability was detected in ggml-org llama.cpp d006858/e15efe0. This affects the function _visit_pattern of the file common/json-schema-to-grammar.cpp. The manipulation results in null pointer dereference. The attack can be launched remotely. The pull request to fix this issue awaits acceptance.

In the Linux kernel, the following vulnerability has been resolved: ipv4: igmp: remove multicast group from hash table on device destruction When a device is destroyed under RTNL, ip_mc_destroy_dev() iterates through the multicast list and calls ip_ma_put() on each membership, scheduling them for RCU reclamation. However, they are not unlinked from the device's multicast hash table (mc_hash). Since the device remains published in dev->ip_ptr until after ip_mc_destroy_dev() completes, concurrent RCU readers traversing mc_hash can still locate and access the multicast group after its refcount is decremented. If the RCU callback runs and frees the group while a reader is accessing it, a use-after-free occurs. Fix this by unlinking the multicast group from mc_hash using ip_mc_hash_remove() before scheduling it for reclamation. BUG: KASAN: slab-use-after-free in ip_check_mc_rcu+0x149/0x3f0 Read of size 4 at addr ffff888009bf1408 by task mausezahn/2276 Call Trace: <IRQ> dump_stack_lvl+0x67/0x90 print_report+0x175/0x7c0 kasan_report+0x147/0x180 ip_check_mc_rcu+0x149/0x3f0 udp_v4_early_demux+0x36d/0x12d0 ip_rcv_finish_core+0xb8b/0x1390 ip_rcv_finish+0x54/0x120 NF_HOOK+0x213/0x2b0 __netif_receive_skb+0x126/0x340 process_backlog+0x4f2/0xf00 __napi_poll+0x92/0x2c0 net_rx_action+0x583/0xc60 handle_softirqs+0x236/0x7f0 do_softirq+0x57/0x80 </IRQ> Allocated by task 2239: kasan_save_track+0x3e/0x80 __kasan_kmalloc+0x72/0x90 ____ip_mc_inc_group+0x31a/0xa40 __ip_mc_join_group+0x334/0x3f0 do_ip_setsockopt+0x16fa/0x2010 ip_setsockopt+0x3f/0x90 do_sock_setsockopt+0x1ad/0x300 Freed by task 0: kasan_save_track+0x3e/0x80 kasan_save_free_info+0x40/0x50 __kasan_slab_free+0x3a/0x60 __rcu_free_sheaf_prepare+0xd4/0x220 rcu_free_sheaf+0x36/0x190 rcu_core+0x8d9/0x12f0 handle_softirqs+0x236/0x7f0

Authentication Bypass by Spoofing vulnerability in team-alembic AshAuthentication allows account takeover of local users via OAuth2/OIDC sign-in. AshAuthentication's OAuth2 and OIDC family strategies matched the local user by email address (an upsert on the email field, or a user-defined sign-in filter) rather than by the OpenID Connect iss/sub claim combination. Per OpenID Connect Core §5.7, only iss/sub uniquely and stably identifies an end-user; other claims, including email, MUST NOT be used as unique identifiers. A provider login presenting a victim's email, including an unverified email, a reused email, or an account with email_verified: false, resolved to and signed in as the victim's existing local account. An unauthenticated attacker who can register an account on any accepted OAuth provider with the victim's email (or who benefits from provider-side email reuse or reclamation) obtains the victim's full local privileges. The fix resolves users by the (strategy, sub) identity stored in a user identity resource, and only links a new sub to an existing local account by email when the provider's email_verified claim is trusted (trust_email_verified?). This issue affects ash_authentication from 0.1.0 before 4.14.0 and from 5.0.0-rc.0 before 5.0.0-rc.10.

Open WebUI is a self-hosted artificial intelligence platform designed to operate entirely offline. Prior to 0.8.11, 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. This vulnerability is fixed in 0.8.11.

openwebui

Open WebUI is a self-hosted artificial intelligence platform designed to operate entirely offline. Prior to 0.9.0, the /api/generate, /api/embed, /api/embeddings, and /api/show endpoints accept any model name from the user and forward the request to the Ollama backend without checking whether the user is authorized to access that model. These endpoints only require get_verified_user (any authenticated non-pending user) and validate that the model exists in the full unfiltered model list, but never check AccessGrants.has_access(). This vulnerability is fixed in 0.9.0.

openwebui

Ollama before 0.17.1 contains a heap out-of-bounds read vulnerability in the GGUF model loader. The /api/create endpoint accepts an attacker-supplied GGUF file in which the declared tensor offset and size exceed the file's actual length; during quantization in fs/ggml/gguf.go and server/quantization.go (WriteTo()), the server reads past the allocated heap buffer. The leaked memory contents may include environment variables, API keys, system prompts, and concurrent users' conversation data, and can be exfiltrated by uploading the resulting model artifact through the /api/push endpoint to an attacker-controlled registry. The /api/create and /api/push endpoints have no authentication in the upstream distribution. Default deployments bind to 127.0.0.1, but the documented OLLAMA_HOST=0.0.0.0 configuration is widely used in practice (large public-internet exposure observed).

ollama

Ollama for Windows contains a Remote Code Execution vulnerability in its update mechanism due to improper handling of attacker‑controlled HTTP response headers. When downloading updates, the application constructs local file paths using values derived from HTTP headers without validation. These values are passed directly to filepath.Join, allowing path traversal sequences (../) to be resolved and enabling files to be written outside the intended update staging directory. An attacker who can influence update responses can exploit this flaw to write arbitrary executables to attacker‑chosen locations accessible to the current user, including the Windows Startup directory. This allows execution of arbitrary executables. Critically, when chained with CVE‑2026‑42248 (Missing Signature Verification for Updates), an attacker can deliver malicious payloads that are written to sensitive locations and executed automatically. Because Ollama for Windows performs silent automatic updates and executes staged binaries without user interaction, this results in automatic and persistent code execution without user awareness. Maintainers of this project were notified early about this vulnerability, but didn't respond with the details of vulnerability or vulnerable version range. Versions from 0.12.10 to 0.17.5 were tested and confirmed as vulnerable, other versions were not tested but might also be vulnerable.

ollama · microsoft

Ollama for Windows does not perform integrity or authenticity verification of downloaded update executables. Unlike other platforms, the Windows implementation of the update verification routine unconditionally returns success so no digital signature or trust validation is performed before staging or executing update payloads, enabling attacker‑supplied executables to be accepted and later executed by the application. Critically, Ollama for Windows performs silent automatic updates, so the malicious payload may be installed automatically without user awareness. Maintainers of this project were notified early about this vulnerability, but didn't respond with the details of vulnerability or vulnerable version range. Versions from 0.12.10 to 0.17.5 were tested and confirmed as vulnerable, other versions were not tested but might also be vulnerable.

ollama · microsoft

A security flaw has been discovered in Ollama up to 0.20.2. This affects the function digestToPath of the file x/imagegen/transfer/transfer.go of the component Tensor Model Transfer Handler. The manipulation of the argument digest results in path traversal. The attack may be performed from remote. This attack is characterized by high complexity. The exploitability is reported as difficult. The exploit has been released to the public and may be used for attacks. The vendor was contacted early about this disclosure but did not respond in any way.

ollama

KubeAI is an AI inference operator for kubernetes. Prior to 0.23.2, the ollamaStartupProbeScript() function in internal/modelcontroller/engine_ollama.go constructs a shell command string using fmt.Sprintf with unsanitized model URL components (ref, modelParam). This shell command is executed via bash -c as a Kubernetes startup probe. An attacker who can create or update Model custom resources can inject arbitrary shell commands that execute inside model server pods. This vulnerability is fixed in 0.23.2.

kubeai

A flaw has been found in Ollama up to 0.18.1. This issue affects some unknown processing of the file server/download.go of the component Model Pull API. Executing a manipulation can lead to server-side request forgery. The attack can be launched remotely. The vendor was contacted early about this disclosure but did not respond in any way.

Endian Firewall version 3.3.25 and prior allow authenticated users to execute arbitrary OS commands via the DATE parameter to /cgi-bin/logs_clamav.cgi. The DATE parameter value is used to construct a file path that is passed to a Perl open() call, which allows command injection due to an incomplete regular expression validation.

endian

llama.cpp is an inference of several LLM models in C/C++. Prior to version b8492, the RPC backend's deserialize_tensor() skips all bounds validation when a tensor's buffer field is 0. An unauthenticated attacker can read and write arbitrary process memory via crafted GRAPH_COMPUTE messages. Combined with pointer leaks from ALLOC_BUFFER/BUFFER_GET_BASE, this gives full ASLR bypass and remote code execution. No authentication required, just TCP access to the RPC server port. This issue has been patched in version b8492.

ggml

A flaw was found in Red Hat OpenShift AI (RHOAI) llama-stack-operator. This vulnerability allows unauthorized access to Llama Stack services deployed in other namespaces via direct network requests, because no NetworkPolicy restricts access to the llama-stack service endpoint. As a result, a user in one namespace can access another user’s Llama Stack instance and potentially view or manipulate sensitive data.

redhat

Wallos is an open-source, self-hostable personal subscription tracker. Prior to version 4.7.0, the patch introduced in commit e8a513591 (CVE-2026-30840) added SSRF protection to notification test endpoints but left three additional attack surfaces unprotected: the AI Ollama host parameter, the AI recommendations endpoint, and the notification cron job. An authenticated user can reach internal network services, cloud metadata endpoints (AWS IMDSv1, GCP, Azure IMDS), or localhost-bound services by supplying a crafted URL to any of these endpoints. This issue has been patched in version 4.7.0.

wallosapp

llama.cpp is an inference of several LLM models in C/C++. Prior to b7824, an integer overflow vulnerability in the `ggml_nbytes` function allows an attacker to bypass memory validation by crafting a GGUF file with specific tensor dimensions. This causes `ggml_nbytes` to return a significantly smaller size than required (e.g., 4MB instead of Exabytes), leading to a heap-based buffer overflow when the application subsequently processes the tensor. This vulnerability allows potential Remote Code Execution (RCE) via memory corruption. b7824 contains a fix.

ggml

llama.cpp is an inference of several LLM models in C/C++. Prior to b8146, the gguf_init_from_file_impl() in gguf.cpp is vulnerable to an Integer overflow, leading to an undersized heap allocation. Using the subsequent fread() writes 528+ bytes of attacker-controlled data past the buffer boundary. This is a bypass of a similar bug in the same file - CVE-2025-53630, but the fix overlooked some areas. This vulnerability is fixed in b8146.

ggml

A vulnerability in the HTML Cascading Style Sheets (CSS) module of ClamAV could allow an unauthenticated, remote attacker to cause a denial of service (DoS) condition on an affected device. This vulnerability is due to improper error handling when splitting UTF-8 strings. An attacker could exploit this vulnerability by submitting a crafted HTML file to be scanned by ClamAV on an affected device. A successful exploit could allow the attacker to terminate the scanning process.

Docker Model Runner (DMR) is software used to manage, run, and deploy AI models using Docker. Versions prior to 1.0.16 expose a POST `/engines/_configure` endpoint that accepts arbitrary runtime flags without authentication. These flags are passed directly to the underlying inference server (llama.cpp). By injecting the --log-file flag, an attacker with network access to the Model Runner API can write or overwrite arbitrary files accessible to the Model Runner process. When bundled with Docker Desktop (where Model Runner is enabled by default since version 4.46.0), it is reachable from any default container at model-runner.docker.internal without authentication. In this context, the file overwrite can target the Docker Desktop VM disk (`Docker.raw` ), resulting in the destruction of all containers, images, volumes, and build history. However, in specific configurations and with user interaction, it is possible to convert this vulnerability in a container escape. The issue is fixed in Docker Model Runner 1.0.16. Docker Desktop users should update to 4.61.0 or later, which includes the fixed Model Runner. A workaround is available. For Docker Desktop users, enabling Enhanced Container Isolation (ECI) blocks container access to Model Runner, preventing exploitation. However, if the Docker Model Runner is exposed to localhost over TCP in specific configurations, the vulnerability is still exploitable.

BigBlueButton is an open-source virtual classroom. In versions 3.0.21 and below, the official documentation for "Server Customization" on Support for ClamAV as presentation file scanner contains instructions that leave a BBB server vulnerable for Denial of Service. The flawed command exposes both ports (3310 and 7357) to the internet. A remote attacker can use this to send complex or large documents to clamd and waste server resources, or shutdown the clamd process. The clamd documentation explicitly warns about exposing this port. Enabling ufw (ubuntu firewall) during install does not help, because Docker routes container traffic through the nat table, which is not managed or restricted by ufw. Rules installed by ufw in the filter table have no effect on docker traffic. In addition, the provided example also mounts /var/bigbluebutton with write permissions into the container, which should not be required. Future vulnerabilities in clamd may allow attackers to manipulate files in that folder. Users are unaffected unless they have opted in to follow the extra instructions from BigBlueButton's documentation. This issue has been fixed in version 3.0.22.

bigbluebutton

ClamAV versions prior to 0.103.0-rc contain a vulnerability in function name processing through the ClamBC bytecode interpreter that allows attackers to manipulate bytecode function names. Attackers can exploit the weak input validation in function name encoding to potentially execute malicious bytecode or cause unexpected behavior in the ClamAV engine.