CVE vulnerability database
Loading data…
Loading data…
LibreChat is an enhanced ChatGPT clone that supports multiple AI providers. Prior to 0.8.7, POST /api/auth/2fa/verify-temp in api/server/routes/auth.js calls verify2FAWithTempToken and accepts repeated TOTP or backup-code guesses without a dedicated attempt limiter. An attacker must first possess valid primary credentials and a valid five-minute temp token, but can then make repeated second-factor attempts and obtain a full session if a valid code is guessed. This issue is fixed in version 0.8.7.
LibreChat is an enhanced ChatGPT clone that supports multiple AI providers. From 0.8.5 until 0.8.6, GET /api/user in api/server/controllers/UserController.js getUserController serializes federatedTokens and tokenset to the authenticated browser because an earlier response-sanitization fix was not applied to this endpoint. Browser-context script can obtain OAuth access, ID, and refresh credentials that were intended to remain server-side. This issue is fixed in version 0.8.6.
LibreChat is an enhanced ChatGPT clone that supports multiple AI providers. Prior to 0.8.7, packages/data-schemas/src/methods/file.ts updateFileUsage accepts chat attachment files[].file_id values without restricting the lookup to the authenticated owner. An authenticated attacker who obtains another user's file identifier, including through GET /api/share/{shareId} for a shared conversation, can attach that file to a chat, causing extracted document text to enter the model context. The same operation can remove expiration metadata from the victim file. This issue is fixed in version 0.8.7.
LibreChat is an enhanced ChatGPT clone that supports multiple AI providers. Prior to 0.8.6, client/src/components/Chat/Input/MCPConfigDialog.tsx renders the remote MCP details.description value through dangerouslySetInnerHTML without sanitization, allowing a malicious MCP server to execute script when a user opens the configuration dialog. The secondary client/src/components/Banners/Banner.tsx sink similarly renders administrator-controlled banner.message HTML without sanitization for all users. Script execution occurs in the LibreChat browser origin and can expose browser-context data or alter displayed content. This issue is fixed in version 0.8.6.
LibreChat is an enhanced ChatGPT clone that supports multiple AI providers. From 0.7.0 until 0.8.6, when local registration is enabled and at least one social provider is configured, POST /api/auth/register passes an unparsed request object into api/server/services/AuthService.js, allowing an unauthenticated attacker to persist a forged provider value for a known victim email. The api/strategies/socialLogin.js email-fallback path then accepts the planted record when the victim signs in because it compares the stored provider name without requiring the provider subject identifier. The victim OAuth session and the attacker local-password session become bound to the same user record, giving the attacker persistent read and write access to data created by the victim. The first-user administrator role is an additional conditional impact only on a fresh deployment. This issue is fixed in version 0.8.6.
LibreChat is an enhanced ChatGPT clone that supports multiple AI providers. Prior to 0.8.6, the POST /api/auth/login local-login validation path in api/strategies/localStrategy.js logs req.body, including the submitted plaintext password, during a Zod validation failure. The standard browser form validation makes this path less likely through the normal interface, but a direct invalid login request can reach it, and the logging formatter does not redact the password stored in metadata. A principal with access to application stdout, error logs, or journal output can recover the submitted credential. This issue is fixed in version 0.8.6.
A weakness has been identified in continuedev continue 2.0.0/2.1.0. Affected is the function validateSearchAndReplaceFilepath of the file gui/src/util/clientTools/editImpl.ts of the component Edit Tool. Executing a manipulation can lead to path traversal. The attack requires local access. The vendor was contacted early about this disclosure but did not respond in any way.
codebase-memory-mcp provides local code intelligence for developers and AI agents through a persistent graph and MCP tools. Prior to 0.10.0, the indexer's .gitignore matcher implemented * and ** through recursive backtracking without a work bound. A repository containing a pattern with repeated wildcard groups and a long path could make matching work grow exponentially, allowing a single ignore rule to stall discovery indefinitely and, before worker isolation, stall the server. The impact is limited to availability, and the bounded matcher does not claim that every expensive input is eliminated. This issue is fixed in version 0.10.0.
codebase-memory-mcp provides local code intelligence for developers and AI agents through a persistent graph and MCP tools. Prior to 0.9.0, automatic import of a repository's .codebase-memory/ artifact allocated a decompression buffer using the attacker-controlled 64-bit original_size value from artifact.json but truncated that value to int when passing the destination capacity to the Zstandard decompressor. A value with bit 31 set became negative and was sign-extended to a large size_t, allowing the decompressor to write frame content beyond the actual allocation. Indexing a malicious repository containing the crafted artifact could corrupt heap memory with attacker-controlled bytes, causing a reliable crash and potentially code execution. This issue is fixed in version 0.9.0.
U.S Department of the Interior Unstable Slope Management Program allows attackers with valid local application credentials to establish an application session by sending JSON POST requests to the application's login endpoint without a Referrer header or completing the intended Keycloak flow. Fixed on or around 2026-10-19.
codebase-memory-mcp provides local code intelligence for developers and AI agents through a persistent graph and MCP tools. Prior to 0.9.0, the Cypher parser accepted an unbounded number of projection items in RETURN and WITH clauses while projection executors indexed fixed 32-element stack arrays using the parsed count. A query_graph query with more than 32 projection items could therefore write beyond those arrays, and a related result-column path could read uninitialized stack slots even when no node matched. Any caller of query_graph, including an agent whose query is influenced by untrusted content, could cause a reliable server crash and potentially compromise process confidentiality and integrity. This issue is fixed in version 0.9.0.
U.S Department of the Interior Unstable Slope Management Program stores local user passwords as unsalted MD5 hashes. Fixed on or around 2026-10-19.
U.S Department of the Interior Unstable Slope Management Program contains a vulnerable endpoint that allows an unauthenticated attacker, without a cookie header, to send a POST request resulting in the creation of a new user. Fixed on or around 2026-10-19.
U.S Department of the Interior Unstable Slope Management Program accepts a POST request from an unauthenticated user, with no cookie header, to a specific endpoint allowing for the deletion of specified users. Fixed on or around 2026-10-19.
codebase-memory-mcp provides local code intelligence for developers and AI agents through a persistent graph and MCP tools. Prior to 0.9.0, the detect_changes tool validated the base_branch argument for shell metacharacters but omitted the leading-dash check before placing the value in the revision position of a git diff command. Git therefore interpreted an option-shaped base_branch value, including --output, as a command option instead of a revision, allowing a caller to create or overwrite a caller-chosen file with Git diff output under the server process's permissions. Arbitrary code execution was not demonstrated. This issue is fixed in version 0.9.0.
U.S Department of the Interior Unstable Slope Management Program allows an unauthenticated user to inject unauthorized SQL commands that allow the attacker to gain root access and complete control over the database. Fixed on or around 2026-10-19.
U.S Department of the Interior Unstable Slope Management Program allows a remote, unauthenticated attacker to access a vulnerable endpoint that returns a list of application users and their permissions in plaintext. Fixed on or around 2026-10-19.
codebase-memory-mcp provides local code intelligence for developers and AI agents through a persistent graph and MCP tools. Prior to 0.9.1-rc.1, the optional graph UI's POST /api/process-kill route omitted the indexing-job ownership check on Windows while retaining that check on POSIX systems. A requester able to reach the optional UI could supply an arbitrary process identifier, causing TerminateProcess to stop any process that the server user was permitted to terminate. The flaw affects the Windows-specific process termination path and can cause a high availability impact without disclosing or modifying data. This issue is fixed in version 0.9.1-rc.1.
MaxKB is an open-source AI assistant for enterprise. Prior to 2.10.7-lts, custom Tool Init Param definitions accept attacker-controlled props_info.rules[].validator strings through POST /admin/api/workspace/{workspace_id}/tool and PUT /admin/api/workspace/{workspace_id}/tool/{tool_id}, store them without restriction, and pass them to eval() when the parameter form renders. An authenticated user with the default USER workspace role can store JavaScript and trigger execution by opening the tool's Debug screen before entering or submitting a parameter value. The same Init Param form is exposed to accounts with edit permission on the tool, but execution by a different account was not independently verified. The confirmed script runs in the MaxKB origin; access to another user's bearer token and authenticated account authority is plausible only if that user renders the malicious form. This issue is fixed in version 2.10.7-lts.
Open Robotics RViz fillIndexBuffer Heap-based Buffer Overflow Remote Code Execution Vulnerability. This vulnerability allows network-adjacent attackers to execute arbitrary code on affected installations of Open Robotics RViz. Authentication is not required to exploit this vulnerability. The specific flaw exists within the handling of mesh resources in PointCloud and Marker displays. The issue results from the lack of proper validation of the length of user-supplied data prior to copying it to a heap-based buffer. An attacker can leverage this vulnerability to execute code in the context of the service account. Was ZDI-CAN-32652.