CVEs

MLflow is an open source AI engineering platform for agents, large language models, and machine learning models. From 3.12.0rc0 until 3.17.0, the scorer invoke, register, and GenAI evaluate endpoints, including unauthenticated POST /ajax-api/3.0/mlflow/scorer/invoke, accept serialized third-party scorer data that Scorer.model_validate_json resolves using a module-path allowlist without constraining the selected class or its constructor arguments. When optional trulens and trulens-providers-litellm packages are installed, an attacker can select mlflow.genai.scorers.trulens, choose a model URI outside MLflow's native provider registry so execution uses the LiteLLM fallback, and supply api_base and api_key values that cause the server to send an outbound request to an attacker-chosen or internal host with attacker-controlled headers and body. The vulnerable deserialization occurs before a tracking-store lookup or job submission. This issue is fixed in version 3.17.0.

The LangGraph CLI is the official command-line interface for LangGraph. From 0.4.15 until 0.4.32, langgraph deploy --verbose passes a temporary registry push token on standard input to docker login --password-stdin, but the CLI subprocess helper prints command input in verbose mode. Anyone who can read resulting CI logs or terminal recordings before the token's maximum 12-hour expiration can pull deployment images or push images to that deployment's image repository. Deployments without --verbose and remote image builds are not affected, and the token alone cannot update a deployment to execute a pushed image. This issue is fixed in version 0.4.32.

LangGraph checkpoint savers provide Postgres- and SQLite-backed state persistence for LangGraph workflows and agents. Prior to langgraph-checkpoint-postgres 3.1.3 and langgraph-checkpoint-sqlite 3.1.2, PostgresStore, AsyncPostgresStore, SqliteStore, and AsyncSqliteStore flatten namespace labels with ".".join(namespace), while BaseStore validates labels only for put and aput. Unvalidated dotted labels passed through get, search, delete, list_namespaces, their asynchronous equivalents, or direct batch and abatch operations can resolve to another tenant's stored namespace, allowing cross-namespace reads, overwrites, or deletion. InMemoryStore is not affected because it compares namespace tuples. This issue is fixed in langgraph-checkpoint-postgres 3.1.3 and langgraph-checkpoint-sqlite 3.1.2.

A security flaw has been discovered in Kakao khaiii up to 0.4. The impacted element is the function khaiii::nn::Linear::open of the file src/main/cpp/khaiii/nn/Linear.cpp of the component JSON Config File Handler. The manipulation of the argument hidden_dim/class_num results in out-of-bounds read. The attack requires a local approach. The exploit has been released to the public and may be used for attacks. The vendor explains: "khaiii treats the resource directory (--rsc-dir) as trusted input, like a binary or a shared library. To use this defect, an attacker must already be able to write to that directory. With that access, an attacker can change the analysis results without this defect." This vulnerability only affects products that are no longer supported by the maintainer.

Twisted is an event-based framework for internet applications, supporting Python 3.6+. Prior to 26.10.0rc1, twisted.names.dns.DNSDatagramProtocol accepts a UDP DNS response for a pending query when its 16-bit transaction ID matches but does not verify that the response came from the queried host and port. An attacker able to deliver a packet to the resolver's listening port with the correct transaction ID can inject a forged answer. The standard twisted.names.client.Resolver uses a new random source port for each query, increasing off-path attack complexity, while applications that reuse a DNSDatagramProtocol instance expose only the transaction-ID search space after the first query. A successful forgery can redirect hostname resolution and enable mail interception, server-side request forgery, or man-in-the-middle positioning. This issue is fixed in version 26.10.0rc1.

OpenReception's appointment booking software provides an end-to-end encrypted appointment booking platform. In versions 2.1.0 and prior, an unauthenticated attacker can store an arbitrary ML-KEM-768 public key under any existing user's identity in any tenant's staff-crypto recipient directory. The product's design guarantees that appointment data is end-to-end encrypted to staff members and that the staff recipient directory is not attacker-writable, so that even administrators and the operator cannot read patient data. This endpoint breaks that guarantee for an unauthenticated remote actor: it allows writing an arbitrary encryption recipient into the recipient directory of any tenant. The injected key receives the tunnel key of every appointment booked afterward in that tenant, because booking clients encrypt to all active staff keys. The attacker holds the corresponding secret key and is a persistent, unauthorized recipient of that key material. At time of publication, there are no publicly available patches.

OpenReception's appointment booking software provides an end-to-end encrypted appointment booking platform. Prior to version 2.2.2, an attacker overbooks an agent, placing multiple appointments into a single time slot that the system is designed to keep exclusive. The number of overlapping appointments equals the number of concurrent requests, and one booking token is enough because the token is consumed after the insert. This results in the integrity of the booking-calendar breaking and availability being denied to other clients. This issue has been patched in version 2.2.2.

OpenReception's appointment booking software provides an end-to-end encrypted appointment booking platform. In versions 2.1.0 and prior, a returning client's six-digit booking PIN can be brute-forced online in the victim tenant. A successful guess yields the client's booking access token, granting access to that client's appointment tunnel, which holds their end-to-end-encrypted appointment data, and the ability to act as the client, for example adding or cancelling appointments in the tunnel. The bypass is cross-tenant: state owned by one tenant removes an authentication protection of a different tenant. This crossing of the tenant boundary is the basis for the scope-changed vector. At time of publication, there are no publicly available patches.

OpenReception's appointment booking software provides an end-to-end encrypted appointment booking platform. Prior to version 2.2.1, an attacker who captures a pending invite token can replay it to obtain a passkey-registration credential and register their own passkey on the invited account, taking control of that account. For a staff or tenant-admin invite this yields access to the tenant context tied to that account. This issue has been patched in version 2.2.1.

OpenReception's appointment booking software provides an end-to-end encrypted appointment booking platform. Prior to version 2.0.0, the account-confirmation link in emails is built from the request Host header, with no allowlist. The unauthenticated POST /api/auth/resend-confirmation rotates a pending invitee's confirmation token and emails a link whose host comes from the request. An attacker who sends one resend-confirmation with Host: attacker.example and a pending invitee's email causes that invitee to receive a confirmation email whose only link points to the attacker. If the invitee clicks, the one-time confirmation token reaches the attacker, who then confirms the account on the real tenant domain and registers a passkey they control, taking over the account before the legitimate invitee ever uses it. All released versions through v1.1.0 are affected; no patched version exists yet. Reaching the app with a spoofed Host depends on the front proxy: the stock Caddy catch-all forwards it, while an operator who configures explicit per-tenant proxy blocks is not affected (see Details 4). The invite path builds links from the trusted tenant.domain (invite/+server.ts:204-210); the resend-confirmation path builds them from the request Host. This issue has been patched in version 2.0.0.