Blog

Overview
Two unsafe pickle deserialization vulnerabilities have been discovered in the SGLang open-source project, one within the tool’s multimodal generation module and another within the Encoder Parallel Disaggregation system. SGLang is a serving framework for large language models (LLMs) and multimodal models. These vulnerabilities, tracked as CVE-2026-3059 and CVE-2026-3060, allow an attacker to provide a malicious pickle file to a device using SGLang’s multimodal generation system or the encoder parallel disaggregation system, enabling code execution. CVE-2026-3989 allows an attacker to provide a malicious pickle file that, when attempted to be replayed with the replay_request_dump.py script, will result in code execution. It is recommended that project maintainers avoid implementing Pickle functions due to the inherent security risks.
Description
SGLang is a framework for serving various large language models (LLMs) and multimodal AI models, supporting models such as Qwen, DeepSeek, Mistral, and Skywork, and is compatible with OpenAI APIs. Two unsafe pickle deserialization vulnerabilities have been identified in the project, tracked as CVE-2026-3059 and CVE-2026-3060.
CVE-2026-3059
SGLangs multimodal generation module is vulnerable to unauthenticated remote code execution through the ZMQ broker, which deserializes untrusted data using pickle.loads() without authentication.
CVE-2026-3060
SGLangs encoder parallel disaggregation system is vulnerable to unauthenticated remote code execution through the disaggregation module, which deserializes untrusted data using pickle.loads() without authentication.
SGLang is vulnerable to CVE-2026-3059 when the multimodal generation system is enabled, and to CVE-2026-3060 when the encoder parallel disaggregation system is enabled. If either condition is met and an attacker knows the TCP port on which the ZMQ broker is listening and can send requests to the server, they can exploit the vulnerability by sending a malicious pickle file to the broker, which will then deserialize it.
CVE-2026-3989
SGLangs replay_request_dump.py contains an insecure pickle.load() without validation and proper deserialization. An attacker can take advantage of this by providing a malicious .pkl file, which will execute the attackers code on the device running the script.
The SGLang project’s replay_request_dump.py script uses pickle.load() without trust validation, allowing for arbitrary code execution if an attacker can control the pickle file contents. This vulnerability has low applicability but high impact, and can be exploited if an attacker can provide a malicious pickle file or write to the crash dump directory, potentially through social engineering or by compromising a directory where crash dump information is automatically saved. The script, intended to replay crash dump information, poses a risk if an attacker can manipulate the input files or directories used by the script.
The use of Pickle is strongly discouraged due to its inherent security risks. Deserializing a pickle file with pickle.loads() can lead to Remote Code Execution (RCE) if an attacker can provide a malicious file. This is because the pickle file format stores not only data but also instructions on how to reconstruct the object, which are executed during deserialization. As a result, an attacker can potentially execute arbitrary Python code.
To mitigate these risks, projects should consider implementing safer serialization formats, such as JSON or XML, or using tools like msgpack to perform more data-driven serialization and deserialization instead of open-ended capabilities such as pickle. This can help prevent RCE attacks and ensure the secure exchange of data.
Impact
An attacker who can send crafted messages to the ZeroMQ interface may trigger unsafe pickle deserialization in SGLang when certain features are enabled, potentially leading to remote code execution. Additionally, an attacker who can provide a malicious pickle file in the context of a crash dump may also exploit CVE-2026-3989. Successful exploitation could allow arbitrary code execution in the context of the SGLang service, potentially leading to host compromise, lateral movement, data exfiltration, or denial-of-service (DoS) attacks. Deployments that expose the affected interface to untrusted networks are at the highest risk of exploitation.
Solution
Users of SGLang should restrict access to the service interfaces and ensure they are not exposed to untrusted networks. Proper network segmentation and access controls should be implemented to prevent unauthorized interaction with the ZeroMQ endpoints. During coordinated disclosure, no vendor statement was obtained. A proposed patch was submitted to the maintainers, but no response was received at the time of publication.
Acknowledgements
Thanks to the reporter, Igor Stepansky. Thank you to Gregory Bowers for additional research assistance during coordination.This document was written by Christopher Cullen.

Overview
Malformed ZIP headers can be used to obfuscate malicious content in ZIP files from antivirus detection tools. Despite the presence of malformed headers, custom extraction software can decompress the ZIP archive, allowing potentially malicious payloads to be recovered after successful antivirus evasion.
After a review of the researcher’s published PoC and further analysis of the case, we have determined that this issue does not meet the necessary criteria to be considered a vulnerability. The originally reported behavior was based on a misunderstanding of native archive-handling capabilities in common operating systems such as Windows and Linux. It was initially believed that native extraction tools could successfully decompress ZIP files with malformed headers, while EDRs and antivirus security tools would fail. After additional testing, we have determined that extraction is not possible through standard archiving tools, and recovering the payload would instead require a custom loader. Accordingly, this technique can be more accurately described as an obfuscation method, as it does not meaningfully violate any implicit or explicit security boundaries.
Description
ZIP archives contain metadata such as compression method, flags, and version information. Antivirus engines typically rely on this metadata to determine how to preprocess files before scanning. If an attacker modifies the compression-method header field, antivirus software may fail to properly decompress the file and will, therefore, be unable to analyze the actual payload.
After antivirus evasion, the payload can be recovered by using a custom loader that ignores the declared Method field and instead decompresses embedded data directly. This allows an attacker to hide malicious content from antivirus engines while still being able to recover it programmatically.
Notably, standard extraction tools (e.g. 7‑Zip, unzip, bsdtar, Python’s zipfile) trust the declared compression method and attempt decompression but then fail with CRC or “unsupported method” errors. These tools do not extract the payload and do not expose the underlying data.
This vulnerability is similar to VU#968818, CVE-2004-0935 but requires unique unpacking tools to enable any abuse or misuse.
Impact
This ZIP header obfuscation method requires custom tooling to fully recover the concealed malicious content. The impact of this obfuscation technique is limited to systems where other security policies (e.g., running arbitrary unzipping tools) may need to be bypassed to enable such an attack.
Solution
Antivirus and EDR vendors should consider alerting on such behavior to ensure their customers are aware of malformed headers in zip content. Users should not use third-party tools or custom unzip capabilities to allow malicious content to be installed on their systems.
Acknowledgements
Thanks to the reporter, Christopher Aziz. This document was written by Laurie Tyzenhaus.

Overview
The Rakuten Viber messaging app for Android V25.7.2.0g and Windows V25.6.0.0-V25.8.1.0, has a flaw in its TLS handshake implementation when using the Cloak proxy configuration. This flaw allows for easy identification of proxy usage, potentially compromising user anonymity.
Description
Rakuten Viber can be configured to use a proxy server. Multiple proxy configurations, including Cloak, are supported. Rakuten Viber in Android V25.7.2.0g and Windows V25.6.0.0–V25.8.1.0 exhibits a flaw in the TLS handshake implementation for Cloak mode. Cloak mode is designed to hide the fact that a proxy or VPN is in use.
However, the Cloak proxy mode has a rigid and easily identified fingerprint, making it trivially identifiable by Deep Packet Inspection (DPI) systems. This allows networks to block Viber traffic, undermining the app’s ability to circumvent censorship and potentially leading to denial of service in certain cases.
Impact
The Cloak-mode proxy traffic fails to hide the use of a proxy. The outgoing data is easily identifiable due to the rigid fingerprint and no longer appears to be normal browser TLS behavior. The user has no indication the proxy is not protecting their data.
Solution
Windows users should upgrade to V27.3.0.0 or later, and Android users should upgrade to V27.2.0.0g or later.
For continued support, Windows users can implement automatic updates for Viber.
Acknowledgements
Thanks to the reporter Oleksii Gaienko, an independent security researcher. This document was written by Laurie Tyzenhaus.

Overview
A command injection vulnerability was identified in the MS-Agent framework that can be triggered through unsanitized prompt-derived input. An attacker can craft untrusted input introduced via a chat prompt or other external content sources, resulting in arbitrary command execution on the target system(s) where the MS-Agent framework is deployed. No patch or vendor statement was obtained during the coordination process.
Description
MS-Agent is a lightweight framework that enables agents to perform autonomous task execution and tool invocation. The MS-Agent framework includes several features, including a Shell tool that enables execution of commands on the target operating system to complete agentic actions.
A vulnerability has been identified that allows unsanitized input to be executed through the Shell tool. This occurs because the software does not sufficiently verify and sanitize content before execution. As a result, an attacker can leverage prompt injection techniques to influence the agent into executing unintended shell commands when interacting with attacker-controlled content.
The Shell tool relies on regular expression–based filtering in the check_safe() method, which is intended to restrict unsafe commands. The implemented default denylist can be bypassed, allowing crafted input to evade validation checks and reach the shell execution layer.
The vulnerability is tracked as:
CVE-2026-2256
Command injection vulnerability in ModelScope’s ms-agent allows an attacker to execute arbitrary operating system commands through crafted prompt-derived input.
This vulnerability may be exploited when the agent is instructed to process or retrieve external content, such as analyzing code, summarizing documents, or performing other tasks that involve interacting with attacker-controlled data. If the retrieved or processed content contains malicious command sequences that bypass the check_safe() validation, the agent may forward those commands to the Shell tool for execution.
The use of a regular expression denylist in the check_safe() method is insufficient to prevent command injection. Denylist-based filtering is inherently fragile and can often be bypassed through encoding, command obfuscation, or alternative shell syntax. In this case, the filtering logic can be circumvented, enabling arbitrary command execution within the execution context of the agent process.
Impact
An attacker who successfully exploits this vulnerability may execute arbitrary operating system commands on the target with the privileges of the MS-Agent process. This may allow modification of system files, lateral movement within the environment, establishment of persistence mechanisms, or exfiltration of sensitive data accessible to the agent.
Solution
No statement was provided by the vendor during coordination efforts. Users should deploy MS-Agent only in environments where ingested content is trusted, validated, or sanitized. Agents with shell execution capabilities should be sandboxed or executed with least-privilege permissions. Additional mitigation strategies include replacing denylist-based filtering with strict allowlists and implementing stronger isolation boundaries for tool execution.
Acknowledgements
Thanks to the reporter, Itamar Yochpaz, for this report. Document written by Christopher Cullen.