CVE Database

The CrewAI CodeInterpreter tool falls back to SandboxPython when it cannot reach Docker, which can enable RCE through arbitrary C function calling.

A command injection vulnerability exists in MLflow's model serving container initialization code, specifically in the `_install_model_dependencies_to_env()` function. When deploying a model with `env_manager=LOCAL`, MLflow reads dependency specifications from the model artifact's `python_env.yaml` file and directly interpolates them into a shell command without sanitization. This allows an attacker to supply a malicious model artifact and achieve arbitrary command execution on systems that deploy the model. The vulnerability affects versions 3.8.0 and is fixed in version 3.8.2.

A path traversal vulnerability exists in the `extract_archive_to_dir` function within the `mlflow/pyfunc/dbconnect_artifact_cache.py` file of the mlflow/mlflow repository. This vulnerability, present in versions before v3.7.0, arises due to the lack of validation of tar member paths during extraction. An attacker with control over the tar.gz file can exploit this issue to overwrite arbitrary files or gain elevated privileges, potentially escaping the sandbox directory in multi-tenant or shared cluster environments.

Perl versions from 5.9.4 before 5.40.4-RC1, from 5.41.0 before 5.42.2-RC1, from 5.43.0 before 5.43.9 contain a vulnerable version of Compress::Raw::Zlib. Compress::Raw::Zlib is included in the Perl package as a dual-life core module, and is vulnerable to CVE-2026-3381 due to a vendored version of zlib which has several vulnerabilities, including CVE-2026-27171. The bundled Compress::Raw::Zlib was updated to version 2.221 in Perl blead commit c75ae9cc164205e1b6d6dbd57bd2c65c8593fe94.

A vulnerability in parisneo/lollms, up to and including version 2.2.0, allows unauthenticated users to upload and process files through the `/api/files/extract-text` endpoint. This endpoint does not enforce authentication, unlike other file-related endpoints, and lacks the `Depends(get_current_active_user)` dependency. This issue can lead to denial of service (DoS) through resource exhaustion, information disclosure, and violation of the application's documented security policies.

OpenClaw before 2026.3.13 allows bootstrap setup codes to be replayed during device pairing verification in src/infra/device-bootstrap.ts. Attackers can verify a valid bootstrap code multiple times before approval to escalate pending pairing scopes, including privilege escalation to operator.admin.

OpenClaw before 2026.3.12 contains a weak authorization vulnerability in Zalouser allowlist mode that matches mutable group display names instead of stable group identifiers. Attackers can create groups with identical names to allowlisted groups to bypass channel authorization and route messages from unintended groups to the agent.

OpenClaw before 2026.3.11 contains an exec allowlist bypass vulnerability where matchesExecAllowlistPattern improperly normalizes patterns with lowercasing and glob matching that overmatches on POSIX paths. Attackers can exploit the ? wildcard matching across path segments to execute commands or paths not intended by operators.

OpenClaw before 2026.3.12 contains an authorization bypass vulnerability where Feishu reaction events with omitted chat_type are misclassified as p2p conversations instead of group chats. Attackers can exploit this misclassification to bypass groupAllowFrom and requireMention protections in group chat reaction-derived events.

OpenClaw before 2026.3.11 contains a privilege escalation vulnerability in device.token.rotate that allows callers with operator.pairing scope to mint tokens with broader scopes by failing to constrain newly minted scopes to the caller's current scope set. Attackers can obtain operator.admin tokens for paired devices and achieve remote code execution on connected nodes via system.run or gain unauthorized gateway-admin access.

GRID::Machine versions through 0.127 for Perl allows arbitrary code execution via unsafe deserialization. GRID::Machine provides Remote Procedure Calls (RPC) over SSH for Perl. The client connects to remote hosts to execute code on them. A compromised or malicious remote host can execute arbitrary code back on the client through unsafe deserialization in the RPC protocol. read_operation() in lib/GRID/Machine/Message.pm deserialises values from the remote side using eval() $arg .= '$VAR1'; my $val = eval "no strict; $arg"; # line 40-41 $arg is raw bytes from the protocol pipe. A compromised remote host can embed arbitrary perl in the Dumper-formatted response: $VAR1 = do { system("..."); }; This executes on the client silently on every RPC call, as the return values remain correct. This functionality is by design but the trust requirement for the remote host is not documented in the distribution.

HTTP::Session versions before 0.54 for Perl defaults to using insecurely generated session ids. HTTP::Session defaults to using HTTP::Session::ID::SHA1 to generate session ids using a SHA-1 hash seeded with the built-in rand function, the high resolution epoch time, and the PID. The PID will come from a small set of numbers, and the epoch time may be guessed, if it is not leaked from the HTTP Date header. The built-in rand function is unsuitable for cryptographic usage. The distribution includes HTTP::session::ID::MD5 which contains a similar flaw, but uses the MD5 hash instead.

Amon2 versions before 6.17 for Perl use an insecure random_string implementation for security functions. In versions 6.06 through 6.16, the random_string function will attempt to read bytes from the /dev/urandom device, but if that is unavailable then it generates bytes by concatenating a SHA-1 hash seeded with the built-in rand() function, the PID, and the high resolution epoch time. The PID will come from a small set of numbers, and the epoch time may be guessed, if it is not leaked from the HTTP Date header. The built-in rand function is unsuitable for cryptographic usage. Before version 6.06, there was no fallback when /dev/urandom was not available. Before version 6.04, the random_string function used the built-in rand() function to generate a mixed-case alphanumeric string. This function may be used for generating session ids, generating secrets for signing or encrypting cookie session data and generating tokens used for Cross Site Request Forgery (CSRF) protection.

Crashmail 1.6 contains a stack-based buffer overflow vulnerability that allows remote attackers to execute arbitrary code by sending malicious input to the application. Attackers can craft payloads with ROP chains to achieve code execution in the application context, with failed attempts potentially causing denial of service.

EChat Server 3.1 contains a buffer overflow vulnerability in the chat.ghp endpoint that allows remote attackers to execute arbitrary code by supplying an oversized username parameter. Attackers can send a GET request to chat.ghp with a malicious username value containing shellcode and ROP gadgets to achieve code execution in the application context.

Bochs 2.6-5 contains a stack-based buffer overflow vulnerability that allows attackers to execute arbitrary code by supplying an oversized input string to the application. Attackers can craft a malicious payload with 1200 bytes of padding followed by a return-oriented programming chain to overwrite the instruction pointer and execute shell commands with application privileges.

MAWK 1.3.3-17 and prior contains a stack-based buffer overflow vulnerability that allows attackers to execute arbitrary code by exploiting inadequate boundary checks on user-supplied input. Attackers can craft malicious input that overflows the stack buffer and execute a return-oriented programming chain to spawn a shell with application privileges.

JAD Java Decompiler 1.5.8e-1kali1 and prior contains a stack-based buffer overflow vulnerability that allows attackers to execute arbitrary code by supplying overly long input that exceeds buffer boundaries. Attackers can craft malicious input passed to the jad command to overflow the stack and execute a return-oriented programming chain that spawns a shell.

TiEmu 2.08 and prior contains a stack-based buffer overflow vulnerability that allows attackers to execute arbitrary code by exploiting inadequate boundary checks on user-supplied input. Attackers can trigger the overflow through command-line arguments passed to the application, leveraging ROP gadgets to bypass protections and execute shellcode in the application context.

JAD 1.5.8e-1kali1 and prior contains a stack-based buffer overflow vulnerability that allows attackers to execute arbitrary code by supplying oversized input that exceeds buffer boundaries. Attackers can craft malicious input strings exceeding 8150 bytes to overflow the stack, overwrite return addresses, and execute shellcode in the application context.

Use of Hard-coded Credentials vulnerability in Microchip Time Provider 4100 allows Malicious Manual Software Update.This issue affects Time Provider 4100: before 2.5.0.

Locutus brings stdlibs of other programming languages to JavaScript for educational purposes. Starting in version 2.0.39 and prior to version 3.0.25, a prototype pollution vulnerability exists in the `parse_str` function of the npm package locutus. An attacker can pollute `Object.prototype` by overriding `RegExp.prototype.test` and then passing a crafted query string to `parse_str`, bypassing the prototype pollution guard. This vulnerability stems from an incomplete fix for CVE-2026-25521. The CVE-2026-25521 patch replaced the `String.prototype.includes()`-based guard with a `RegExp.prototype.test()`-based guard. However, `RegExp.prototype.test` is itself a writable prototype method that can be overridden, making the new guard bypassable in the same way as the original — trading one hijackable built-in for another. Version 3.0.25 contains an updated fix.

Locutus brings stdlibs of other programming languages to JavaScript for educational purposes. Prior to version 3.0.25, the `unserialize()` function in `locutus/php/var/unserialize` assigns deserialized keys to plain objects via bracket notation without filtering the `__proto__` key. When a PHP serialized payload contains `__proto__` as an array or object key, JavaScript's `__proto__` setter is invoked, replacing the deserialized object's prototype with attacker-controlled content. This enables property injection, for...in propagation of injected properties, and denial of service via built-in method override. This is distinct from the previously reported prototype pollution in `parse_str` (GHSA-f98m-q3hr-p5wq, GHSA-rxrv-835q-v5mh) — `unserialize` is a different function with no mitigation applied. Version 3.0.25 patches the issue.

Notesnook is a note-taking app. Prior to version 3.3.11 on Web/Desktop and 3.3.17 on Android/iOS, a stored XSS in the Web Clipper rendering flow can be escalated to remote code execution in the desktop app. The root cause is that the clipper preserves attacker-controlled attributes from the source page’s root element and stores them inside web-clip HTML. When the clip is later opened, Notesnook renders that HTML into a same-origin, unsandboxed iframe using `contentDocument.write(...)`. Event-handler attributes such as `onload`, `onclick`, or `onmouseover` execute in the Notesnook origin. In the desktop app, this becomes RCE because Electron is configured with `nodeIntegration: true` and `contextIsolation: false`. Version 3.3.11 Web/Desktop and 3.3.17 on Android/iOS patch the issue.

Handlebars provides the power necessary to let users build semantic templates. In versions 4.0.0 through 4.7.8, `Handlebars.compile()` accepts a pre-parsed AST object in addition to a template string. The `value` field of a `NumberLiteral` AST node is emitted directly into the generated JavaScript without quoting or sanitization. An attacker who can supply a crafted AST to `compile()` can therefore inject and execute arbitrary JavaScript, leading to Remote Code Execution on the server. Version 4.7.9 fixes the issue. Some workarounds are available. Validate input type before calling `Handlebars.compile()`; ensure the argument is always a `string`, never a plain object or JSON-deserialized value. Use the Handlebars runtime-only build (`handlebars/runtime`) on the server if templates are pre-compiled at build time; `compile()` will be unavailable.

Federated Learning and Interoperability Platform (FLIP) is an open-source platform for federated training and evaluation of medical imaging AI models across healthcare institutions. The FLIP login page in versions 0.1.1 and prior has no rate limiting or CAPTCHA, enabling brute-force and credential-stuffing attacks. FLIP users are external to the organization, increasing credential reuse risk. As of time of publication, it is unclear if a patch is available.

Gematik Authenticator securely authenticates users for login to digital health applications. Versions prior to 4.16.0 are vulnerable to authentication flow hijacking, potentially allowing attackers to authenticate with the identities of victim users who click on a malicious deep link. Update Gematik Authenticator to version 4.16.0 or greater to receive a patch. There are no known workarounds.

Langflow is a tool for building and deploying AI-powered agents and workflows. Prior to version 1.9.0, the Agentic Assistant feature in Langflow executes LLM-generated Python code during its validation phase. Although this phase appears intended to validate generated component code, the implementation reaches dynamic execution sinks and instantiates the generated class server-side. In deployments where an attacker can access the Agentic Assistant feature and influence the model output, this can result in arbitrary server-side Python execution. Version 1.9.0 fixes the issue.

Home Assistant is open source home automation software that puts local control and privacy first. Home Assistant apps (formerly add-ons) configured with host network mode expose unauthenticated endpoints bound to the internal Docker bridge interface to the local network. On Linux, this configuration does not restrict access to the app as intended, allowing any device on the same network to reach these endpoints without authentication. Home Assistant Supervisor 2026.03.02 addresses the issue.

Pi-hole Admin Interface is a web interface for managing Pi-hole, a network-level ad and internet tracker blocking application. Versions prior to 6.0 have a critical OS Command Injection vulnerability in the savesettings.php file. The application takes the user-controlled $_POST['webtheme'] parameter and concatenates it directly into a system command executed via PHP's exec() function. Since the input is neither sanitized nor validated before being passed to the shell, an attacker can append arbitrary system commands to the intended pihole command. Furthermore, because the command is executed with sudo privileges, the injected commands will run with elevated (likely root) privileges. Version 6.0 patches the issue.