Currently browsing: Technology

Overview
The Kaltura HTML5 Player V2 Library (mwEmbed / html5lib) contains two vulnerabilities, both involving the same insecure deserialization flaw, that enable arbitrary file read and remote code execution. Affected versions include html5lib v2.45, v2.103 and earlier, and other v2.x releases that expose the vulnerable mwEmbedLoader.php endpoint. Notably, only versions of the legacy player (Player V2) are vulnerable; these issues do not affect any versions of the currently supported Kaltura Player V7.
Description
Kaltura is an AI video platform that provides tools for video management, publishing, playback, and integration with web applications. Kaltura’s HTML5 player library exposes the mwEmbedLoader.php endpoint, which accepts a user-controlled ServiceUrl parameter as the target URL for backend API requests. The KalturaClientBase PHP client library fetches data from this URL and automatically deserializes it using PHP’s unserialize() function without validating source, scheme, or content.
CVE-2026-19913 results from the combination of this unsafe deserialization flaw and improper error-handling behavior. An attacker can provide the location of a local file to ServiceUrl as a file:// path, and the client will fetch the internal file’s contents and attempt to deserialize them. When deserialization fails, the raw bytes are reflected back to the client in the resulting error message, enabling the attacker to read any file accessible to the web-server user.
CVE-2026-19912 is caused by insufficient sanitization of the parameter uiconf_id, which is appended to the base cache folder path when the application writes data to disk. Because this value is user-controlled and unsanitized, an attacker can supply values that include directory traversal sequences such as ../ to redirect file writes outside the intended cache directory. When the deployment uses the default file-based cache backend, an attacker can direct ServiceUrl to a malicious serialized object containing executable PHP code, then supply a uiconf_id path value that writes its deserialized fields to a web-accessible directory. The attacker can then request the file directly to achieve remote code execution as the web-server user. A memcache-only backend may suppress the file write and prevent this specific code-execution path, but the underlying unsafe deserialization behavior and unsanitized path construction remain present.
Impact
These vulnerabilities allow a remote, unauthenticated attacker to read arbitrary local files and execute arbitrary commands as the web-server user. No authentication or Kaltura session token is required to exploit either issue; an attacker only needs network access to the affected html5lib endpoint.
CVE-2026-19913 can be abused to obtain database credentials, administrative secrets, API keys, or any other sensitive information hosted on the affected instance. Remote code execution achieved through CVE-2026-19912 allows an attacker to modify or exfiltrate platform data, deploy tools for persistence and lateral movement, and further compromise affected Kaltura deployments. Because the affected endpoint is also exposed on Kaltura’s shared, multi-tenant CDN infrastructure, these vulnerabilities affect not only individual customer installations, but also every tenant served by these shared hosts.
Solution
Kaltura has released new patches to remediate these vulnerabilities in all affected legacy Player V2 versions. Customers using legacy players, including self-hosted legacy player deployments (html5lib v2.x), should update to the patched version or, preferably, migrate to the newer and currently supported Kaltura Player V7 platform. Until the update is applied, users are advised to restrict or disable access to the mwEmbedLoader.php endpoint.
Acknowledgements
Thanks to Gerjan Wemekamp (AndDone) for researching and reporting these vulnerabilities. This document was written by Molly Jaconski.

Overview
Konami’s Metal Gear Online 3 video game contains a heap-based buffer overflow that can be triggered by an input‑validation vulnerability that allows match hosts to remotely execute arbitrary code on lobby members’ machines through specially crafted data.
Description
Metal Gear Online 3 is an online 8 vs. 8 competitive shooter game that uses Steam Matchmaking to handle its multiplayer lobbies and matches. As detailed in CVE‑2026‑19874, version 1.1.2.8 of Metal Gear Online 3 (Steam AppID 287700) contains an input‑validation vulnerability in the processing of Steam lobby metadata related to the player‑removal feature. The game tracks a lobby field (kick_num) indicating the number of players designated for removal, along with corresponding Steam ID entries of each kicked player (kicked_id_%i). When joining a lobby, the player’s client parses these fields to check whether its own Steam ID is on the list, and if so, the player is prevented from joining the match.
The function responsible for parsing this lobby data does not validate the kick_num value against the size of the fixed‑length buffer allocated for kicked player identifiers. Supplying a kick_num value larger than the buffer capacity results in out‑of‑bounds writes into adjacent memory. The memory region immediately following this buffer contains internal Steamworks callback handler structures that store function pointers and callback arguments for processing lobby data changes, messages, and other related events. By manipulating the overflow, an attacker can corrupt these handler structures and redirect callback execution, resulting in control‑flow hijacking on affected client systems. The vulnerability can be triggered automatically when a client joins a lobby controlled by an attacker.
Impact
Exploitation of this vulnerability may allow remote code execution on affected clients. Initial control‑flow hijacking provides access only to existing in‑process code; however, the Metal Gear Online 3 binary includes Denuvo‑protected regions mapped with read‑write‑execute (RWX) permissions. These regions permit runtime injection of attacker‑supplied code, significantly increasing the severity of the issue. An attacker hosting a lobby can achieve code execution on any client that joins, without requiring further interaction from the victim. Additionally, because host privileges are automatically reassigned to another lobby participant when the current host exits, an attacker can obtain host control during an active match and subsequently deliver the malicious lobby data to all connected players. This enables compromise of multiple systems through a single exploitation event.
Solution
As of this writing, Konami has not released patch notes or an advisory that specifically addresses this vulnerability, but a fix was included in version 1.1.2.9 of the Metal Gear Online 3 executable, mgsvmgo.exe. The patch also iterated the server and lobby version numbers from 15 to 16 and 150 to 160, respectively, to prevent players on older versions from accessing the online services.
This is the latest patch that fixed the vulnerability:
https://steamdb.info/patchnotes/24176213/
The full patch list can be found here:
https://steamdb.info/app/287700/patchnotes/
Acknowledgements
Thank you to Alice Cecchetto for reporting this vulnerability. This document was written by Bob Kemerer.

Overview
The Calix GS7 XGS GS5239XG router running firmware EXOS/6.6.47 contains a missing authentication vulnerability that exposes its UPnP (Universal Plug and Play) WANIPConnection service on the public WAN interface.
Description
Calix GS7 XGS GS5239XG is a residential gateway that provides routing, NAT, and firewall functionality for home networks. The device includes the Universal Plug and Play (UPnP) service implemented via MiniUPnPd 2.3.7, a lightweight software program that provides features such as automatic port forwarding for applications and devices on the LAN. By default, the UPnP service is exposed on the device’s WAN interface and does not require authentication.
CVE-2026-75501 In affected firmware versions, the router binds its UPnP WANIPConnection SOAP service to the public WAN interface on TCP port 5000. Because the service does not require authentication when accepting SOAP requests, a remote attacker can obtain full access to the router’s critical UPnP functions including adding, deleting, and enumerating NAT port mappings.
Impact
CVE-2026-75501 enables an unauthenticated, remote attacker to remotely query and manipulate existing NAT mappings. By exploiting this vulnerability to create arbitrary port-forwarding rules on the router, an attacker can bypass NAT and firewall protections, exposing internal LAN devices to the public internet. Because the Calix router is typically provisioned with its default UPnP-enabled configuration, this issue poses significant risk to residential users with network-connected internal devices such as security cameras, network-attached storage (NAS), and other IoT appliances.
Solution
Unfortunately, the CERT/CC was unable to reach Calix to coordinate this vulnerability. Until a vendor patch is available, users can reduce exposure by disabling UPnP on the router’s administrative interface. If the UPnP setting is unavailable or locked, it may be necessary to contact your ISP to request its deactivation at the carrier level. Alternatively, filtering inbound traffic to TCP port 5000, either via the router itself, a secondary firewall, or through your ISP, can prevent external hosts from reaching the WANIPConnection service.
Acknowledgements
Thanks to Brian Khan Quintana for researching and reporting this vulnerability. This document was written by Molly Jaconski.

Overview
RDK Central RDK-B WebUI version, rdkb-2025q4-kirkstone, contains multiple vulnerabilities involving memory corruption, improper authentication, race conditions, and insufficient input validation. An attacker with network access to an affected WebUI may be able to bypass authentication, obtain administrative access, cause a denial-of-service condition, or corrupt memory within underlying RDK-B processes. Under certain conditions, this memory corruption may potentially be leveraged for arbitrary code execution.
Description
RDK-B (Reference Design Kit for Broadband) is an open-source software platform used in broadband gateways and related networking devices. The RDK-B WebUI provides a web-based interface for configuring and administering an RDK-B device. Five vulnerabilities have been identified in the RDK-B WebUI.
CVE-2026-19505JWT (JSON Web Token) authentication in javascript-templates/source/jst_functions.c does not correctly verify whether a token’s cryptographic signature is valid. The application treats both a valid signature and an invalid signature as successful verification because it incorrectly checks the return value from OpenSSL’s EVP_VerifyFinal() function.
A remote, unauthenticated attacker can craft a JWT with an invalid signature that is still accepted by the WebUI. Successful exploitation allows the attacker to log in as the privileged user and gain administrative access to the device.
CVE-2026-19506 The login process in /usr/www2/check.jst uses a shared value to store the result of password verification. Because this value is shared between multiple requests, the application may return one user’s authentication result to another user’s session.
An unauthenticated attacker can send a login request at the same time a legitimate administrator logs in. If the requests are timed correctly, the attacker’s session may receive the administrator’s successful authentication result, allowing access to the WebUI without knowing the correct password.
CVE-2026-19507 The login handler in /usr/www2/check.jst does not limit the length of the password submitted by a user. The application performs SHA-256 hashing on the entire supplied password before rejecting the login attempt.
A remote, unauthenticated attacker can submit very large password values to consume excessive CPU resources. Repeated requests can make the WebUI and related services slow or unresponsive, resulting in a denial-of-service condition.
CVE-2026-19508 The data parser in javascript-templates/source/jst_post.c does not properly validate malformed input before processing it in memory. A remote, unauthenticated attacker can send a specially crafted request that causes the Duktape WebUI (https://duktape.org/) process to access or modify memory incorrectly.
During data parser processing and later during Duktape memory cleanup, indicating that application memory can be corrupted. An attacker may be able to use this vulnerability to cause a denial-of-service and potentially execute arbitrary code, although code execution has not been demonstrated.
CVE-2026-19509 The ajaxSet_wireless_network_configuration.jst handler does not properly validate the ssid_number value before passing it to the RDK-B routing service.
An authenticated administrator can supply an abnormally large value that causes memory corruption in the native rtrouted process. This condition can crash rtrouted and trigger an RBus service restart. Successful exploitation can cause a denial-of-service. Because the flaw results in native memory corruption, arbitrary code execution may also be possible.
Impact
A remote, unauthenticated attacker with network access to the RDK-B WebUI may be able to bypass authentication and obtain administrative access to the device. An unauthenticated attacker may also cause memory corruption or resource exhaustion, resulting in denial-of-service.
An authenticated administrator may be able to trigger memory corruption in the privileged rtrouted process and result in arbitrary code execution.
Successful exploitation of one or more of these vulnerabilities could result in unauthorized administrative access, modification of device configuration, loss of availability, or potentially execution of attacker-controlled code on an affected device.
Solution
Unfortunately, RDK Central was unreachable to coordinate these vulnerabilities. Until an update is available, administrators should restrict access to the RDK-B WebUI to trusted management networks and authorized hosts. The administrative interface should not be exposed directly to the Internet or other untrusted networks.
Acknowledgements
Thanks to Mikołaj Pisula and Michał Bernacki for researching and reporting these vulnerabilities. This document was written by Michael Bragg.

Overview
Two vulnerabilities have been identified in the Trusted Platform Module (TPM) 2.0 reference implementation:

CVE-2026-6726 – Information leakage via falsified TPM keys.
CVE-2026-6727 – A timing side-channel vulnerability in RSA OAEP decryption.

An attacker with privileged access to a TPM command interface may be able to exploit these vulnerabilities by sending specially crafted TPM commands. Successful exploitation could allow the attacker to decrypt ciphertexts encrypted to affected TPM-managed RSA keys, including the RSA Endorsement Key (EK), or obtain credentials for falsified TPM keys, enabling forged TPM 2.0 attestations.
These vulnerabilities are also documented by the Trusted Computing Group (TCG) in advisories – TCGVRT010 and TCGVRT0011:
Description
Trusted Platform Module (TPM) technology provides hardware-backed cryptographic services for modern computing platforms. TPMs are designed to resist tampering and may be implemented as discrete chips, integrated hardware, firmware-based TPMs (fTPMs), or software implementations used in cloud and virtualized environments.
The Trusted Computing Group (TCG) maintains the TPM specifications and publishes a reference implementation to assist vendors in developing TPM-compliant products.
Two vulnerabilities were identified in the TPM 2.0 reference implementation.
CVE-2026-6727
A timing side-channel vulnerability exists in the RSA OAEP decryption implementation. A privileged local attacker with access to the TPM command interface may be able to exploit timing differences to recover information that could allow decryption of ciphertexts encrypted to TPM-managed RSA keys, including the RSA Endorsement Key (EK), including import blobs, credential blobs, and session salts. Under certain conditions, this may also enable the forgery of TPM 2.0 attestations.
CVE-2026-6726
An information leakage vulnerability could allow a privileged local attacker to obtain credentials from a TPM-aware Certificate Authority (CA) for a falsified TPM key, such as an Attestation Key (AK), DevID key, or TLS authentication key. This could enable the creation of fraudulent TPM 2.0 attestations using the forged key.
Both vulnerabilities require privileged access to the TPM command interface. Multiple vendors have released firmware and software updates incorporating fixes from the updated TPM 2.0 reference implementation.
Impact
Successful exploitation requires privileged local access to a TPM command interface. Depending on the vulnerability exploited, an attacker may be able to:
– Decrypt ciphertexts encrypted to the TPM-managed RS keys, including the RSA Endorsement Key (EK), , including credential blobs, import blobs, and session salts.
– Obtain credentials for falsified TPM keys.
– Produce fraudulent TPM 2.0 attestations that appear to originate from a legitimate TPM.
The overall impact depends on the affected TPM implementation and how TPM-based attestation and key management are used by the platform.
Solution
The vulnerabilities originate in the TPM 2.0 reference implementation, and TPM vendors have incorporated the corresponding fixes into updated firmware and software releases. Users should install TPM firmware updates, operating system updates, or software patches provided by their platform or TPM vendor.
Cloud providers using software-based TPM implementations may also have deployed updates. Customers should consult their cloud provider’s guidance to determine whether any additional action is required. See the Vendor Information section for product-specific remediation guidance.
Acknowledgements
Thanks to security researchers Liran Perez, Zecharye Galitzky, Shai Sarfati, and Yanai Moyal from Intel for reporting these vulnerabilities. Thanks to members of the Trusted Computing Group’s Vulnerability Response Team, TCG VRT, for working with CERT/CC towards this multi-party vulnerability disclosure. This document was written by Vijay Sarvepalli.

Overview
The OpenCart v4.2.0.0 extension installer contains a directory traversal vulnerability. The extension installation process extracts uploaded .zip files then uses the zip entry filenames as filesystem paths, without validating that the resolved path stays inside the intended directory. This vulnerability is tracked as CVE-2026-18412.
Description
OpenCart is a free, open‑source e‑commerce solution designed to help businesses build and manage online stores.
OpenCart extensions are uploaded as zip files with .ocmod.zip extensions. Upon installation, the OpenCart v4.2.0.0 extension installer extracts these zip files, but does not validate that the extracted paths stay inside the intended extraction directory. An attacker can craft a malicious extension containing file path traversal sequences, such as ../. With this vulnerability, an attacker can write files, such as a PHP web shell, into the webroot directory.
Impact
If a user with valid admininistrator credentials installs a malicious extension, it could allow a user to remotely execute code with the same privileges that OpenCart has on the target server. This includes the potential creation of a web shell, which could further enable remote execution of system-level commands. The vulnerability was confirmed against version 4.2.0.0, but other 4.x versions of OpenCart may be affected.
Solution
Unfortunately, OpenCart could not be reached to coordinate this vulnerability, and a patch is not available at the time of this writing. The CERT/CC recommends that OpenCart users update to the latest version and avoid installing extensions from unknown or untrusted sources. Additionally, OpenCart should be configured to run with the minimum privileges necessary for normal operation.
Acknowledgements
Thank you to Noah Magill for reporting this vulnerability. This document was written by Bob Kemerer.

Overview
A heap buffer overflow vulnerability exists in the stb TrueType library created by nothings. Exploitation of this vulnerability can occur when handling malformed font data and may lead to both Denial of Service (DoS) and Information Disclosure.
Description
The nothings stb repository, versions 1.26 and earlier, contains a collection of single-file public domain and MIT-licensed libraries for C/C++ projects.
CVE-2026-18497 A heap buffer overflow vulnerability exists in the stbtt_GetGlyphShape() function within the stb_truetype.h library when handling malformed TrueType Font (TTF) data. The issue occurs during glyph contour parsing. The function iterates based on the number of contour endpoints specified in endPtsOfContours, but does not validate that the points pointer remains within the bounds of the glyph data buffer. As a result, operations such as flags = *points++; may read memory beyond the allocated region.
An attacker can exploit this vulnerability by crafting a TTF file that specifies an excessively large endPtsOfContours value while providing truncated or minimal glyph data. When an application utilizing stb_truetype.h attempts to load, bake, or render this malformed font via stbtt_GetGlyphShape(), the parser will continue iterating past the end of the glyph data buffer, triggering the out-of-bounds heap read.
Impact
Exploitation of this vulnerability can lead to application crashes due to invalid reads of unmapped memory addresses, resulting in denial of service (DoS). In some cases, an attacker may also be able to read data from adjacent heap memory.
Solution
Unfortunately, we were unable to reach the stb maintainers to coordinate this vulnerability. Users are advised to monitor the project’s GitHub repository for updates and install the latest version of this library once a fix has been released. At the time of this writing, the maintainer of this repository had the following message posted on the project’s README page: “This project discusses security-relevant bugs in public in Github Issues and Pull Requests, and it may take significant time for security fixes to be implemented or merged. If this poses an unreasonable risk to your project, do not use stb libraries.”
Acknowledgements
Thank you to Yanzhao Shen for reporting this vulnerability. This document was written by Bob Kemerer.

Overview
A cross-site scripting (XSS) vulnerability in Alinto SOGo v5.12.7 allows attackers to achieve remote code execution by embedding malicious SVG (Scalable Vector Graphics) objects in ICS (iCalendar) invitations. The vulnerability has been actively exploited in the wild, as confirmed by VirusTotal sightings.
Description
Alinto SOGo is an open-source webmail and groupware platform for email, calendars, contacts, and shared scheduling. It is primarily used by organizations seeking a self-hosted interface solution for existing mail infrastructure.
CVE-2026-8496
The vulnerability exists in SOGo’s handling of ICS files, where the DESCRIPTION field is rendered without proper sanitization or Content Security Policy (CSP) enforcement. When a calendar invite contains an SVG payload, such as <animate onrepeat=’…’>, with JavaScript event handlers, the browser executes the script in the context of the SOGo webmail interface. This occurs during normal calendar view rendering (e.g., when a user opens or previews the calendar tab), even without explicit interaction.
Impact
Once an attacker delivers a malicious ICS file via email, any user who views the calendar will execute the embedded XSS payload, granting the attacker full read access to the victim’s mailbox. This enables:

Credential theft through forced logout/login phishing
Password manager autofill hijacking
Full exfiltration of email messages, folder contents, and metadata
Extraction of contact lists and calendar data

Solution
Users are recommended to upgrade to SOGo v5.12.8 or newer. v5.12.8 addresses this issue via sanitizing ICS DESCRIPTION content and stricter handling of embedded SVG and HTML.
Acknowledgements
Thank you to Greg Lesnewich for reporting this issue. This AI-assisted vulnerability note was prepared by Alexander Curtis.

Overview
VPS.org’s one-click deployment templates provision services with default passwords and predefined network bindings instead of generating randomized secrets or applying per-deployment hardening measures.
Description
VPS.org is a cloud and virtual private server hosting provider that offers a library of templates for quickly provisioning common applications and services. Multiple vulnerabilities exist in the one-click deployment templates feature. These vulnerabilities stem from the same root cause: content is directly instantiated from static templates, using default passwords and static secrets with no deployment-specific randomization or interface-binding hardening at provisioning time.
CVE-2026-16503 The Supabase template provides an instance of PostgreSQL that is bound to all network interfaces (0.0.0.0:5432) and uses the hard-coded database password postgres. Because Docker manages its own iptables rules, this exposure can bypass standard host UFW firewall configurations. If the the instance is exposed to the internet, a remote attacker could connect to the host’s published TCP port 5432 and authenticate as the postgres superuser account using the default postgres password.
CVE-2026-16504 The Zulip template ships with a hard-coded application key secret_key: changeme, a default database password zulip, and the setting DISABLE_HTTPS=True. An attacker can use this public secret key to forge or validate signed session material, enabling session forgery and authentication bypass against the instance. If unchanged, the default database password zulip can be used to authenticate to the database. Furthermore, the DISABLE_HTTPS=True configuration causes all traffic to be sent over unencrypted HTTP by default, exposing credentials and session data to potential interception in certain deployments.
Impact
CVE-2026-16503 (Supabase template): PostgreSQL superuser access from the internet enables the following:
* read and exfiltrate data
* insert/modify/delete data
* alter the database schema, roles, and privileges
* establish persistence via database objects
* denial of service through destructive statements (dropping tables/databases)
CVE-2026-16504 (Zulip template): Authentication bypass and session forgery allows the following:
* account and instance takeover
* interception of credentials and session tokens over unencrypted transport
This constitutes a Technical Impact = Total under the SSVC framework, meaning:

The vulnerability gives the adversary total control over the behavior of the software or total disclosure of all information on the affected system.

Solution
Unfortunately, VPS.org could not be reached to coordinate these vulnerabilities, and a patch is not yet available. Users of VPS.org one-click deployment templates are advised to change any default passwords and secret keys before deploying to a production environment. Firewall rules and network segmentation should be implemented to restrict internet access to back-end systems such as databases and other sensitive resources. Finally, where applicable, HTTPS should be enabled to protect credentials and session data in transit.
Acknowledgements
Thank you to Simon Gajdosik for reporting these vulnerabilities. This document was written by Bob Kemerer.

Overview
Six vulnerabilities have been discovered within the SGLang project, including remote code execution (RCE), server-side request forgery (SSRF), local file read, credential leakage, and model weight exfiltration on a target server. Exploitation does not require authentication in most cases, and some vulnerabilities require only network access with no API keys or user credentials. At the time of publication, no patches are available from the project maintainers, and coordination attempts have been unsuccessful.
Description
SGLang is an open-source framework for serving large language models (LLMs) and multimodal AI models, supporting models such as Qwen, DeepSeek, Mistral, and Skywork, and is compatible with OpenAI APIs. Six vulnerabilities have been discovered within the tool and are tracked as follows:
CVE-2026-15969
SGLang contains unauthenticated RCE in /load_lora_adapter_from_tensors by bypass of SafeUnpickler’s incomplete denylist, allowing arbitrary command execution through crafted base64-encoded pickle payloads.
CVE-2026-15971
SGLang contains an RCE vulnerability when the optional dumper subsystem is enabled, which allows for a sandbox escape when DUMPER_SERVER_PORT is set, enabling code execution on inference requests.
CVE-2026-15974
SGLang contains an SSRF and local file read in the multimodal generation endpoint /v1/chat/completions because image_url input is unsanitized, allowing access to internal metadata, secrets, and services.
CVE-2026-15976
SGLang contains a RCE vulnerability when attempting to load model weights from a HuggingFace repository, specifically within the /update_weights_from_disk, where torch.load(…, weights_only=False) fallback enables pickle deserialization of .bin files.
CVE-2026-15977
SGLang contains a credential leakage vulnerability in the /server_info endpoint, which returns API keys and SSL keyfile information when only the –admin-api-key is configured.
CVE-2026-15978
SGLang contains a model weight exfiltration vulnerability when no API keys are configured, because SGLang will expose two endpoints that allow a remote attacker to trigger distributed weight broadcasting using NCCL and then triggering data transfer, attackers can exfiltrate all model weights.
Impact
If exploited, these vulnerabilities could allow an unauthenticated attacker to achieve remote code execution, exfiltrate model weights, or overwrite arbitrary files on the host machine running SGLang. Deployments that expose the affected interface to untrusted networks are at the highest risk of exploitation.
Solution
Until a patch is available, affected users should consider the following mitigations:
Mitigations

Restrict access to the service interfaces and ensure they are not exposed to untrusted networks.
Implement network segmentation and access controls to prevent unauthorized interaction with the vulnerable endpoints.
Change SGLANG_USE_PICKLE_IPC to “false” within environ.py.
Disable endpoints not in use to remove potential attack vectors.

The SGLang maintainers have begun addressing pickle deserialization vulnerabilities and are working to refactor the code base with msgpack to prevent deserialization issues such as CVE-2026-14890, but the SGLANG_USE_PICKLE_IPC defaults to true within the codebase at the time of writing.
Acknowledgements
Thanks to the reporter, Apoorv Dayal [apoorvdayal@outlook.com]. This document was written by Christopher Cullen.