The internet is broken right now.

Altair is a GraphQL client for all platforms. From 7.2.4 until 8.5.8, packages/altair-iframe-sandbox/src/main.ts reads plugin_sandbox_opts from the iframe-sandbox query string and passes attacker-controlled scriptUrls to injectScript() in packages/altair-core/src/utils/inject.ts, which creates script elements in the page. The intended parent iframe flow places the page in an iframe with sandbox allow-scripts and without allow-same-origin, but direct navigation to a crafted iframe-sandbox URL omits that isolation and executes the supplied script in the origin serving Altair. Exploitation requires a victim to open the crafted URL, and the resulting script can access script-readable storage and make same-origin requests according to the deployment's authentication and cookie controls. This issue is fixed in version 8.5.8.

XSS

russh is a Rust SSH client and server library. From version 0.34.0-beta.1 until version 0.64.0, its SSH server does not require the signature algorithm in a public-key authentication signature to match the algorithm named in the authentication request. A client possessing an authorized RSA private key can announce  rsa-sha2-512  while supplying a valid  ssh-rsa  SHA-1 signature, bypassing restrictions or producing misleading audit records when applications rely on the requested algorithm. A valid private-key signature remains required, and russh clients are not affected. This issue is fixed in version 0.64.0.

Crypto Library

Angular is a development platform for building mobile and desktop web applications using TypeScript/JavaScript and other languages. Prior to 20.3.34 and 21.0.3, the @angular/common HttpClient function xsrfInterceptorFn uses a regular expression that does not normalize request URLs in the same way as the browser's WHATWG URL parser. With XSRF protection enabled, a mutating HTTP request to an attacker-controlled URL containing backslashes, stripped control characters or spaces, or removable ASCII whitespace can be misclassified as relative and same-origin, causing the interceptor to attach the X-XSRF-TOKEN header before the browser dispatches the request to a cross-origin host. With permissive CORS handling for the custom header, the attacker can obtain the victim's XSRF token and use it to perform cross-site request forgery against the victim session. This issue is fixed in versions 20.3.34 and 21.0.3.

Other Browser & mail

The Model Context Protocol TypeScript SDK provides TypeScript client and server implementations of the Model Context Protocol. From 1.6.0 until 1.32.0 in @modelcontextprotocol/sdk, from 2.0.0 until 2.3.0 in @modelcontextprotocol/server, from 2.0.0 until 2.0.2 in @modelcontextprotocol/express, and from 2.0.0 until 2.3.1 in @modelcontextprotocol/server-legacy, server-side bearer authentication could accept a valid access token issued for a different service because the SDK did not enforce the token audience. The affected surface is server-side HTTP bearer authentication. The trigger is a valid token whose audience names another service. The authorization server must issue tokens for multiple services, and the application's verifyAccessToken() must not independently validate aud. Applications already validating aud inside verifyAccessToken() are not affected. An attacker with a token for another service could authenticate to the MCP server and obtain the confidentiality and integrity capabilities granted to an authenticated caller. Upgrading alone leaves the new audience validation disabled. Operators must enable audience validation and provide the token audience together. Enabling only the expected audience check causes every request to receive HTTP 401. This issue is fixed in @modelcontextprotocol/sdk 1.32.0, @modelcontextprotocol/server 2.3.0, @modelcontextprotocol/express 2.0.2, and @modelcontextprotocol/server-legacy 2.3.1.

Auth & access Library

Pillow is a Python imaging library. In versions 2.3.0 through 12.3.0, EPS rendering does not validate the scaled image dimensions before invoking Ghostscript. Applications that process untrusted EPS or PostScript files using an application-selected scale greater than one can incur excessive Ghostscript CPU and temporary disk consumption. In versions with decompression-bomb protection, the scaled output is checked only after Ghostscript has completed rendering, allowing resource consumption beyond the configured pixel limit. The amplification depends on the scale selected by the application, not an attacker-controlled scale value. This issue is fixed in version 13.0.0.

Denial of service Library

Pillow is a Python imaging library. In versions through 12.3.0, the libtiff TIFF encoder can pass arguments to  TIFFSetField()  that do not match libtiff’s definition of the corresponding tag. When an application saves attacker-controlled image metadata through this encoder, crafted tag values can cause invalid memory reads, copying process memory into the encoded output or crashing the process. Applications that return the encoded output to an attacker can disclose that memory. The issue is reachable when re-saving compressed TIFF images and through certain conversions that preserve EXIF metadata into compressed TIFF output. This vulnerability is fixed in version 13.0.0.

Denial of service Library

Black is the uncompromising Python code formatter. Prior to 26.10.1, the Black GitHub Action with use_pyproject: true accepts tool.black.required-version from pyproject.toml without validating that it is a released Black version specifier. An attacker who can submit a pull request can change that field to a direct reference to attacker-controlled code, causing the workflow to install and execute the code in the victim repository's GitHub Actions context. This issue is fixed in version 26.10.1.

Other

GraphQL Tools provides utilities for building, stitching, and mocking GraphQL schemas. From 10.10.1 until 12.0.2, the @graphql-tools/utils mergeIncrementalResult helper passes incrementalResult.path segments to setObjectKeyPath without runtime primitive-type validation. A non-primitive segment can bypass strict comparisons against __proto__, constructor, and prototype, then be coerced during property access so a malicious or compromised upstream incremental GraphQL response writes attacker-controlled properties to Object.prototype in the consuming Node.js process. The affected implementation is packages/utils/src/mergeIncrementalResult.ts and is reachable when consumers merge untrusted @defer or @stream transport results. This issue is fixed in version 12.0.2.

Other

Improper Handling of Highly Compressed Data (Data Amplification) vulnerability in silviucpp ezstd allows an attacker who supplies a zstd frame to ezstd:decompress_using_ddict/2 to exhaust memory and terminate the Erlang VM via a small frame that declares a very large content size. zstd_nif_decompress_using_ddict in c_src/ezstd_nif.cc reads the frame's declared content size with ZSTD_getFrameContentSize and passes it directly to enif_make_new_binary, without the 1 GiB MAX_BUFFER_SIZE check that ezstd:decompress/1 applies before allocating. A valid 65,550-byte frame of RLE blocks declaring 2 GiB is rejected by ezstd:decompress/1 but is fully allocated and decompressed by the dictionary path. The allocation happens inside a NIF, so when it exceeds the memory available to the VM the operating system kills the whole node rather than the calling process. This issue affects ezstd: from 1.0.5 before 1.2.5.

Other

Wails is a framework for building desktop applications using Go and web technologies. From 2.0.0 until 2.13.0, BindingsAllowedOrigins in v2/internal/frontend/originvalidator/originValidator.go passes wildcard patterns to wildcardPatternToRegex, which expands every asterisk to an unrestricted regular-expression match. When an application configures a trailing or partial-label wildcard immediately after an allowed domain, an attacker-controlled lookalike origin can match the allowed suffix and invoke bound Go methods through IPC. A complete-component subdomain wildcard is an unaffected configuration. The resulting confidentiality, integrity, and availability impact depends on the file, database, command, credential, or other operations exposed by the application's bound methods. This issue is fixed in version 2.13.0.

Other Server

dcmjs is a JavaScript implementation of DICOM manipulation. Prior to 0.52.0, the DecimalString and IntegerString implementations in src/ValueRepresentation.js sanitize text with a permissive regular expression and pass the result to JavaScript's Number constructor without validating finiteness or the DICOM value-representation rules. Scientific notation overflow in a DS or IS field can produce positive or negative Infinity, while stripping alphabetic characters from signed infinity text can leave a sign-only string that converts to NaN. IntegerString also accepts exponent notation and values outside the DICOM signed 32-bit IS range. Applications that trust these parsed values can propagate invalid numbers into spatial geometry, display windowing, dose scaling, frame counts, rescaling, or volume reconstruction, causing incorrect calculations or disrupted processing. This issue is fixed in version 0.52.0.

Other

MONAI (Medical Open Network for AI) is an AI toolkit for health care imaging. Prior to 1.6.1, PydicomReader._get_affine() trusts non-finite values in the DICOM PixelSpacing, ImagePositionPatient, and ImageOrientationPatient fields and composes them into an affine matrix without validation. An unauthenticated attacker who can upload a crafted DICOM to a default MONAILabel deployment can propagate NaN or infinity values through LoadImaged, Spacingd, model inference, and output writing, causing inference failures, persistent datastore poisoning, or geometrically corrupted results. This issue is fixed in version 1.6.1.

Other

OAuthLib is an implementation of the OAuth request-signing logic for Python 3.6+. From 3.3.0 until 4.0.1, OAuthLib's RFC 8628 device workflow does not require DeviceCodeGrant.validate_token_request to check that the device_code parameter is present and valid. A caller whose client request is accepted by the configured validator and who knows a valid client ID enabled for the device workflow can submit a missing or bogus device code to the token endpoint and obtain a bearer token without completing device authorization. This issue is fixed in version 4.0.1.

Auth & access

Emissary is a P2P based data-driven workflow engine. Prior to 8.51.0, the VALID_LINK validation in src/main/java/emissary/server/mvc/NavAction.java accepts protocol-relative URLs when processing configured navigation links. Browser URL normalization can interpret these accepted strings as external origins. The authoritative advisory attributes open redirect, server-side request forgery, and cloud metadata access to this bypass, but it identifies neither the server-side dereference path nor how navigation-link processing reaches that path. This issue is fixed in version 8.51.0.

Disclosure Browser & mail

Caddy is an extensible server platform that uses TLS by default. Prior to 2.11.5, the reverse proxy's finalizeResponse method in modules/caddyhttp/reverseproxy/reverseproxy.go handles an upstream 101 Switching Protocols response before running removeConnectionHeaders and the hopHeaders deletion loop used for ordinary responses. The early return invokes handleUpgradeResponse, which copies the upstream headers to the client and normalizes only WebSocket-specific header names. An attacker who controls or compromises a proxied WebSocket upstream can therefore cause hop-by-hop headers such as Alt-Svc, Keep-Alive, and Proxy-Authenticate, as well as headers named by Connection, to cross the proxy boundary. For HTTP/2 or HTTP/3 extended CONNECT, the upstream 101 response may be represented to the client as an ordinary successful response, increasing the chance that leaked headers are processed. This issue is fixed in version 2.11.5.

Other Server

borgbackup is an opensource, deduplicating archiver with compression and authenticated encryption. Prior to 1.4.5 and 2.0.0b22, extracting a malicious archive can follow a symlinked parent path or resolve embedded dot-dot path components and delete or overwrite files outside the extraction directory. A crafted slave hardlink can also use an attacker-controlled hardlink source outside the extraction directory and hard-link that file into the restored tree, disclosing its contents. Creating the malicious archive requires repository access and, for encrypted or authenticated repositories, the Borg key and passphrase; a victim must then extract the archive. This issue is fixed in versions 1.4.5 and 2.0.0b22.

Other Device

Remark42 is a self-hosted comment engine for blogs, articles, or any other place where readers can add comments. From 1.0.0 until 1.16.2, Remark42 globally trusts True-Client-IP, X-Real-IP, and X-Forwarded-For without verifying that the direct peer is a trusted proxy, and middleware.RealIP overwrites r.RemoteAddr before per-IP controls run. In a directly reachable deployment, or a proxy that forwards attacker-controlled headers, a remote attacker can rotate spoofed forwarding headers to obtain fresh token buckets for rate-limited endpoints, bypass vote de-duplication when --votes-ip is enabled, mint anonymous identities, and poison stored comment IPs and request logs. The global concurrent-request throttle and per-request timeouts still limit availability impact. Mitigation support is available in version 1.16.2, but operators must configure --trusted-proxy for the deployment topology.

Other

Sliver is a command and control framework. Prior to 1.7.6, an HTTP(S) listener using the PNG encoder accepts attacker-controlled PNG dimensions without an effective decoded-size or pixel limit in util/encoders/images.go, while the listener nonce space implemented in server/encoders/encoders.go is small enough to search remotely. An unauthenticated attacker who identifies a valid nonce can submit a highly compressed image to PNGEncoder.Decode, where it expands during png.Decode, exhausting server memory and causing a complete loss of availability. This vulnerability is fixed in 1.7.6.

Denial of service Server

KurrentDB is a database engineered for modern software applications and event-driven architectures. In version 26.1.0, a user authorized to read the $all stream can use Secondary Indexing query capabilities to read arbitrary files accessible to the database process, enumerate directories, and issue HTTP GET requests from that process. The default configuration limits this privilege to administrators. The flaw does not permit file writes and does not expose environment variables. This issue is fixed in version 26.1.1.

Other Server

Cronicle is a multi-server task scheduler and runner, with a web based front-end UI. Prior to 0.9.124, an authenticated non-admin with run_events, create_events, or edit_events privilege can supply undeclared keys in event params through api_create_event, api_update_event, or api_run_event. lib/api.js and lib/api/event.js validate the parameter object without restricting its keys, and lib/job.js uppercases those keys and applies them to the child process environment after administrator-defined plugin variables. A key that collides with a security-sensitive environment variable can therefore override the plugin execution context and may cause privilege escalation, arbitrary code execution, data access, or service disruption depending on the plugin and runtime. This issue is fixed in version 0.9.124.

Code execution Web app

Code Graph RAG is an AI-powered knowledge graph and retrieval system for querying, understanding, and editing multi-language codebases. Prior to 0.0.589, AstGrepService in codebase_rag/tools/ast_grep_service.py enumerates files beneath project_root and relies on should_skip_path in codebase_rag/utils/path_utils.py to classify them, but the containment check operates on unresolved lexical paths. A symbolic link planted inside an analyzed repository can therefore reference a file outside the configured root while still appearing in scope. The structural_search tool follows the link and discloses readable content without an approval gate, while structural_replace with dry_run=False follows it and overwrites writable content; the MCP path invokes structural replacement without the agent-loop approval control. An attacker who causes a victim to analyze a repository containing such a link can read credentials, keys, environment files, or sibling source and can modify files accessible to the code-graph-rag process. This issue is fixed in version 0.0.589.

Disclosure

The Model Context Protocol TypeScript SDK provides TypeScript client and server implementations of the Model Context Protocol. From 2.0.0 until 2.3.0 in @modelcontextprotocol/server, applications that shared one stateless HTTP transport or one McpServer or Server instance among multiple clients could deliver responses and notifications to the wrong client. The affected surface is an MCP server reached over HTTP by multiple clients. The trigger is reuse of one stateless transport or server instance. Messages were associated with a connection using only the JSON-RPC request identifier. SDK clients start request identifiers at zero, so collisions occur in normal operation. Cross-client delivery exposes one client's responses and notifications to another client. The issue requires an HTTP endpoint reachable by more than one client and an application that constructs the server or transport once at startup instead of once per request or session. stdio servers and factory-created per-call servers are not affected. Upgrading causes reused instances to fail and therefore requires an application code change. This issue is fixed in @modelcontextprotocol/server 2.3.0.

Other Library

The Model Context Protocol TypeScript SDK provides TypeScript client and server implementations of the Model Context Protocol. From 1.3.0 until 1.32.0 in @modelcontextprotocol/sdk and from 2.0.0 until 2.3.0 in @modelcontextprotocol/server, HTTP servers did not enforce a limit on the number of array elements and object keys validated in one request. The affected surface is validation of HTTP requests containing arrays or object keys. The trigger is an array-element or object-key validation workload large enough to exhaust memory while remaining within the request size limit. A single request within the default 4 MiB limit can terminate the process with JavaScript heap exhaustion. No credentials are required unless the application authenticates before SDK processing. Tools accepting arrays or records carry the greatest risk. stdio-only servers do not expose the HTTP request path. Upgrading alone does not enable the new validation bound. Operators must configure both a tool-input element bound and an appropriate request body size limit. This issue is fixed in @modelcontextprotocol/sdk 1.32.0 and @modelcontextprotocol/server 2.3.0.

Denial of service Library

The Model Context Protocol TypeScript SDK provides TypeScript client and server implementations of the Model Context Protocol. Prior to 1.32.0 in @modelcontextprotocol/sdk and from 2.0.0 until 2.3.0 in @modelcontextprotocol/client, the HTTP client transports and OAuth client requests followed redirects to a different origin while retaining custom request headers and, for HTTP 307 or 308 responses, request bodies. The affected surfaces are StreamableHTTPClientTransport, SSEClientTransport, and OAuth client functions such as auth(). The trigger is an HTTP redirect to another origin. Custom headers and mcp-session-id can be re-sent to the redirected origin. HTTP 307 and 308 redirects can also re-send the request body. OAuth request bodies can contain refresh_token and client_secret values. Current runtimes already strip the Authorization header across origins. stdio clients do not use the affected HTTP path. A malicious or compromised MCP or authorization server could therefore receive API keys, session identifiers, OAuth refresh tokens, or OAuth client secrets. Restoring unrestricted redirect following after upgrading restores the exposure. This issue is fixed in @modelcontextprotocol/sdk 1.32.0 and @modelcontextprotocol/client 2.3.0.

Disclosure Library

Dolibarr is an open-source enterprise resource planning and customer relationship management application. Prior to 22.0.5 and 23.0.1, there is a SQL injection vulnerability in the Third Party company update functionality. The application concatenates the  localtax1_value  and  localtax2_value  properties into an SQL UPDATE statement without numeric conversion or other appropriate sanitization. An authenticated user with Third Party create or edit permission can submit crafted  lt1  or  lt2  POST parameters to read sensitive database information, including password hashes, API keys, and business data, or modify third-party records. This issue is fixed in versions 22.0.5 and 23.0.2.

Injection Server

Insecure direct object reference (IDOR) vulnerability in the MatchPoint API, which relies on the `actingasuser` HTTP header without verifying the requester's authorization for the specified user. Manipulation of sequential identifiers allows access to personal, banking, and financial data, private messages, and QR codes for physical access, as well as the ability to modify third-party accounts, change passwords, impersonate users, and transfer virtual balances.

Auth & access Device

Credentials are stored in plain text in an SQLite database and in the MatchPoint Android app's SharedPreferences. An attacker with access to the device's local data or a backup could retrieve usernames and passwords and use them to gain unauthorized access to accounts on the platform.

Auth & access Device

SQL injection in the push notification handler of the MatchPoint Android app, which processes messages from Firebase Cloud Messaging and constructs queries by concatenating untrusted data. An attacker capable of interacting with the component could manipulate queries and access sensitive information from the local database, including credentials stored in the versions described in the report. The scope of the injection is limited to the device.

Injection Device

filelock is a platform-independent file lock for Python. From 3.22.0 until 4.2.1, soft lock classes on Windows read an existing marker with _read_lock_file and remove a stale marker through break_lock_file using pathname operations that wait for an oplock break. Another local account with read access to an unheld marker can keep a Level 1 or batch oplock unanswered, causing SoftFileLock.acquire and related async or fallback locks to wait beyond timeout, blocking=False, and cancel_check. A stale marker under the oplock cannot be broken, but mutual exclusion remains intact and no file other than the marker is accessed. This issue is fixed in version 4.2.1.

Denial of service Library

filelock is a platform-independent file lock for Python. From 3.30.0 until 4.2.1, On POSIX platforms, StrictSoftFileLock and AsyncStrictSoftFileLock validate the coordination directory in _ensure_protocol_directory but later reach the claims directory through _open_directory by pathname, allowing a local account that can rename the coordination name to swap in a symbolic link between check and use during acquisition. That race can create and leave claim files in another writable directory or strand claims in a different strict lock so the second lock times out. Separately, the claims property and force_break follow a resting symbolic link without a race when the lock user inspects or recovers the lock. Claims inspection discloses a file name to the other account only if the resulting exception text or claim_name reaches that account through a shared log, CI output, or error report. A blind force_break recovery that removes live claims from another strict lock can allow two processes into that lock's critical section; this path reaches other strict locks, not arbitrary files. This issue is fixed in version 4.2.1.

Other Library

LXC is a userspace interface for the Linux kernel containment features used to create and manage system containers. In LXC 7.0.0 and earlier, lxc-copy and the public liblxc clone() API can follow attacker-controlled symbolic links while updating /etc/hostname in a copied container root filesystem. The affected surfaces are lxc-copy and liblxc's public lxc_container::clone() API. The privileged cloning process must run as root. The source rootfs must be attacker-controlled or less trusted. A symlink can be placed at etc/hostname or in an intermediate component. The clone operation preserves the symlink before updating the hostname. The vulnerable write truncates and overwrites an existing file outside the destination rootfs. When root clones an untrusted container without LXC_CLONE_KEEPNAME, the copied symlink can resolve outside the destination rootfs and cause the privileged process to overwrite an existing host file with the destination container name. Unprivileged use and managers with independent cloning implementations are not affected. LXC_CLONE_KEEPNAME avoids the hostname update and therefore the affected write.

Disclosure OS & kernel

A static credential embedded in the mobile app has been reported to return the 'autorizado' token on the MatchPoint platform. The report states that this token grants unauthenticated access to API services.

Auth & access Device

filelock is a platform-independent file lock for Python. From 3.10.0 until 4.2.1, filelock on Windows ignores the mode argument for WindowsFileLock files and the ReadWriteLock SQLite database. A lock created with mode=0o600 receives the parent directory's inheritable ACL because _open_non_reparse_fd and _nt_open do not create an owner-only security descriptor. Another local account with inherited access can read holder metadata or the lock database, modify those files where the ACL permits writes, or take the byte-range lock after release and delay later acquisitions. UnixFileLock is not affected, and mutual exclusion between cooperating processes continues to hold. This issue is fixed in version 4.2.1.

Disclosure Library

filelock is a platform-independent file lock for Python. From 3.20.4 until 4.2.1, ReadWriteLock and AsyncReadWriteLock on Windows use _connect to attach SQLite to the lock database by pathname, allowing another local account that can create adjacent names to place a database-wal or database-journal side file before acquisition. An undeletable WAL file makes each read or write acquisition fail immediately with sqlite3.OperationalError regardless of timeout. Separately, when the database contains at least one table, a hot journal that grants the lock account modify rights can cause SQLite recovery to delete a file named by that journal only when the lock account has permission to delete the target. The issue requires a directory writable by another account, such as a shared Windows temporary directory, and does not affect private directories. This issue is fixed in version 4.2.1.

Denial of service Library

BTCPay Server before 2.4.2, including 2.4.2 release candidates, contains an information disclosure vulnerability that allows unauthenticated remote attackers to retrieve .macaroon credential files for an LND (Lightning Network Daemon) node. Attackers can use the exposed macaroon to gain full control of the LND node, including moving funds, resulting in theft of funds from affected deployments.

Disclosure Device

Lack of authentication and authorization in the Images.ashx endpoint in MatchPoint, which allows profile photos to be downloaded using predictable identifiers. An unauthenticated remote attacker could enumerate these identifiers and collect users' photos. Linking these photos to identities may be facilitated by other vulnerabilities in the platform.

Auth & access Web app

Incorrect validation of TLS certificates in the MatchPoint Android app due to a TrustManager that accepts any certificate and a permissive hostname validator. Additionally, the app allows unencrypted HTTP traffic. An attacker capable of intercepting the victim's traffic could read and modify communications, including credentials and personal data.

Crypto Device

API keys, credentials, and configuration parameters embedded in the MatchPoint APK, which can be extracted through reverse engineering. The vendor distinguishes the extracted elements from Google Maps and Firebase keys that must remain on the client side and considers them project identifiers, not authentication secrets. The impact depends on the capabilities and restrictions of each exposed element.

Auth & access Device

filelock is a platform-independent file lock for Python. From version 3.30.0 until version 4.2.1,  StrictSoftFileLock  and  AsyncStrictSoftFileLock  on POSIX systems create coordination and claim directories without applying the caller’s restrictive lock mode. A permissive umask or inherited default ACL can leave these directories writable by another local user, allowing that user to delete an active holder’s claims. Another authorized process can then acquire the lock while the original holder remains in its critical section. An attacker can also plant malformed claims to prevent acquisition. This issue is fixed in version 4.2.1.

Other Library

MinIO Signature V4 verification before RELEASE.2026-09-22T19-25-18Z accepts unsigned x-amz-* headers, letting a presigned PUT URL be escalated into a cross-bucket read.

Crypto Server

Rucio is a scientific data management framework. From 1.16.1 until 35.9.0, 38.6.0, and 41.0.0rc1, the /credentials/signurl endpoint reads X-Rucio-Account instead of the account stored in request.environ['issuer'] by token authentication. Any authenticated low-privilege user can supply root as the header value and pass the root-only get_signed_url permission check. The server then returns a presigned read, write, or delete URL for an attacker-selected object using stored S3, GCS, or Swift credentials, and X-Rucio-VO can select another tenant in multi-VO deployments. This permits data disclosure, modification, deletion, and cross-tenant access on cloud-backed RSEs. This issue is fixed in versions 35.9.0, 38.6.0, and 41.0.0rc1.

Auth & access Web app

Kubeflow Pipelines enables users to build and deploy portable, scalable machine learning workflows. Prior to 2.17.0, the API server's pipeline and pipeline-version creation paths fetch a caller-supplied pipelineURL without restricting the destination. An authenticated caller with permission to create a pipeline or pipeline version can make requests from the API server to internal services and observe a limited status-code oracle. Arbitrary response-body disclosure, credential theft, and integrity or availability impact are not established. This issue is fixed in version 2.17.0.

Disclosure Web app

Kubeflow Pipelines enables users to build and deploy portable, scalable machine learning workflows. Prior to 2.17.0, the API server trusts structural SQL identifier fields decoded from the base64-encoded JSON page_token parameter before constructing list queries. An attacker who can invoke an affected pipeline or run list API can inject through KeyFieldName or SortByFieldName, including v1beta1 and v2beta1 run-listing metric paths, and read data available to the API server database account across namespaces. Database modification is not established. This issue is fixed in version 2.17.0.

Injection Web app

Kubeflow Pipelines enables users to build and deploy portable, scalable machine learning workflows. Prior to 2.15.0, the frontend routes /pipeline/k8s/pod and /pipeline/k8s/pod/events use frontend/server/k8s-helper.ts to include untrusted podname and podnamespace query values and backend error details in HTML-capable error responses. An unauthenticated attacker can craft a link that executes JavaScript in the frontend origin when a user who satisfies the deployment's authentication and route-access requirements follows it and reaches an affected response. Cookie disclosure depends on cookie attributes, and server-side code execution or availability impact is not established. This issue is fixed in version 2.15.0.

XSS Web app

Kubeflow Pipelines enables users to build and deploy portable, scalable machine learning workflows. From 2.17.0 until 2.18.0, the TensorBoard proxy token introduced by the incomplete remediation for CVE-2026-54745 uses MINIO_SECRET_KEY as its HMAC secret, whose bundled default is publicly known. In an authorization-disabled deployment using that default, an unauthenticated attacker can forge a token containing a chosen viewerName and namespace; the proxy constructs the destination from the viewerName plus the -service suffix, the namespace, and the configured cluster domain, then reads responses on port 80 under the /tensorboard/<viewer>/ path. Resource-name validation and path-traversal rejection remain enforced, authorization-enabled deployments still check access to the selected namespace, and unrestricted metadata access, writes, or service disruption are not established. This issue is fixed in version 2.18.0.

Auth & access Web app

Kubeflow Pipelines enables users to build and deploy portable, scalable machine learning workflows. Prior to 2.18.0, the frontend GET /artifacts/get route accepts source=http or source=https with caller-controlled bucket and key values, and isAllowedResourceName permits an all-digit decimal host that the URL parser converts to an IPv4 address. The default ALLOWED_ARTIFACT_DOMAIN_REGEX and authorization-disabled standalone configuration allow an unauthenticated caller to fetch internal HTTP resources and receive their response body. In the authenticated multi-user configuration, the same destination-validation defect is reachable by a low-privilege tenant with a valid identity and authorized namespace. Cloud credential disclosure depends on reachable metadata services and attached roles, and the issue does not establish IMDSv2 bypass or credential access when metadata is blocked. This issue is fixed in version 2.18.0.

Disclosure Web app

Kubeflow Pipelines enables users to build and deploy portable, scalable machine learning workflows. Prior to 2.16.1, the multi-user API server treats the empty namespace used by shared or global pipelines as an authorization exception for modification and deletion operations. An authenticated user can modify or delete shared pipeline definitions and versions without the required Kubernetes RBAC permissions, compromising their integrity and availability. Read-only access to shared pipelines remains intentional, single-user deployments intentionally bypass multi-user checks, and execution of a modified pipeline is a conditional downstream consequence. The separate shared-pipeline multipart upload authorization path is outside this advisory and is not remediated by version 2.16.1. This issue is fixed in version 2.16.1.

Auth & access Web app

Kubeflow Pipelines enables users to build and deploy portable, scalable machine learning workflows. Prior to 2.18.0, the multi-user HTTP endpoints POST /apis/v2beta1/pipelines/upload and POST /apis/v2beta1/pipelines/upload_version call canUploadVersionedPipeline, which returns without SubjectAccessReview when the target namespace is empty while REQUIRE_NAMESPACE_FOR_PIPELINES defaults to false. A low-privilege tenant can create shared pipelines or inject attacker-controlled versions into existing shared pipelines without the required write permission; if the API server is directly reachable without an authenticating proxy, an unauthenticated caller can exploit the same bypass, while deployments protected by such a proxy do not expose anonymous access. An enabled unpinned recurring run that follows the latest version can then execute the injected specification in another tenant's namespace under that workload's service account, while pinned or embedded-specification runs are not automatically substituted. This issue is fixed in version 2.18.0.

Auth & access Server

Kubeflow Pipelines enables users to build and deploy portable, scalable machine learning workflows. Prior to 2.18.0, the frontend artifact download path for source=s3 or source=minio accepts providerInfo.Params.endpoint, and the archived pod-log path accepts S3-compatible endpoint metadata from workflow status, without requiring those endpoints to match operator-controlled storage configuration. In the shipped authorization-disabled standalone configuration, an unauthenticated caller can make the shared ml-pipeline-ui pod connect to arbitrary reachable hosts and return response data; requests may also reveal an access-key identifier when signed with configured object-store credentials. Multi-user secret-backed metadata remains restricted to the server namespace. This issue is fixed in version 2.18.0.

Disclosure Server

Kubeflow Pipelines enables users to build and deploy portable, scalable machine learning workflows. Prior to 2.18.0, the legacy V1/raw-Argo execution cache mutating webhook and cache-server storage path hash selected template fields without the Kubernetes namespace, use the key alone for lookups and duplicate checks, and trust the mutable pipelines.kubeflow.org/execution_cache_key annotation when publishing outputs. In multi-user deployments where tenant namespaces share the cache service, a tenant with workload-submission privileges can read another namespace's cached outputs or seed matching outputs that are substituted into a later step. Standard KFP V2 component pods bypass this cache, matching requires the selected template fields and staleness policy, and arbitrary code execution or service availability impact is not established. This issue is fixed in version 2.18.0.

Code execution Server