A security vulnerability has been detected in HKUDS nanobot up to 0.2.1. Impacted is the function ExecTool._guard_command/ExecTool._spawn of the file nanobot/agent/tools/shell.py of the component Shell Allowlist Handler. Such manipulation leads to os command injection. The attack can be executed remotely. The exploit has been disclosed publicly and may be used. Upgrading to version 0.3.0 is recommended to address this issue. The name of the patch is 4562. It is advisable to upgrade the affected component. Multiple issues were reported to the project. They reacted with a high level of professionalism and kindness: "These five reports are variants of the same root cause: validation of shell commands containing multiple segments, wrappers, comments, or chained commands. The issue was fixed by validating every executable shell segment against the configured allowlist".
LightRAG provides simple and fast retrieval-augmented generation. Prior to 1.5.4, when LightRAG is deployed with LIGHTRAG_API_KEY set but AUTH_ACCOUNTS unset, X-API-Key protection can be bypassed because lightrag/api/auth.py falls back to a hardcoded DEFAULT_TOKEN_SECRET, /auth-status and /login can mint guest JWTs, and combined_dependency in lightrag/api/utils_api.py accepts a valid guest token before checking the API key. A remote unauthenticated attacker can call endpoints guarded by combined_auth, including document read, upload, deletion, graph mutation, and query endpoints. This vulnerability is fixed in 1.5.4.
LightRAG provides simple and fast retrieval-augmented generation. Prior to 1.5.4, the server defaults to CORS_ORIGINS=* combined with allow_credentials=True in lightrag/api/lightrag_server.py, causing Starlette CORSMiddleware to effectively whitelist every origin for credentialed cross-origin requests. Any malicious website visited by an authenticated LightRAG user can silently make authenticated API requests, exfiltrating documents and knowledge graph data or performing destructive actions such as deleting the document store. This vulnerability is fixed in 1.5.4.
DeepTutor before version 1.4.10 contains an authorization bypass vulnerability that allows low-privilege users to invoke unrestricted MCP tools due to the allowed_mcp_tools function returning None instead of a denied result when mcp_tools is omitted from a user's grant in deeptutor/multi_user/tool_access.py. Attackers or prompt-injected content acting within a user session can enumerate and invoke any configured MCP tool, including filesystem, shell, and browser servers, gaining unauthorized access to sensitive deployment resources.
Vibe-Trading before 0.1.10 contains a path traversal vulnerability that allows attackers to write files outside the intended memory root directory by supplying a malicious memory_type value containing path traversal sequences through the remember tool. Attackers can manipulate the memory_type parameter in the persistent memory store to cause the application to write arbitrary Markdown files to unintended locations on the filesystem.
Vibe-Trading before 0.1.10 constructs the swarm run directory by joining a caller-supplied run identifier onto the runs base directory without validation in run_dir (agent/src/swarm/store.py). A crafted run identifier supplied through the MCP swarm tools causes the application to read arbitrary run.json files outside the runs directory and to overwrite existing run.json files at traversed locations.
Vibe-Trading before 0.1.10 builds the proposal file path by joining a caller-supplied proposal identifier onto the broker proposals directory without sanitization (agent/src/live/mandate/commit.py). A proposal identifier containing path traversal sequences causes the application to load an attacker-controlled JSON file as an authoritative live trading mandate. Combined with the file upload endpoint, an admitted caller can write a JSON file to a known location and traverse to it, and because the ceilings validation is skipped when ceilings are absent, the attacker fully controls the committed mandate.
Vibe-Trading before 0.1.10 contains a DNS rebinding authentication bypass vulnerability that allows remote attackers to bypass bearer-token authentication by exploiting the server's trust of TCP peer addresses for loopback clients combined with missing Host header validation while binding to 0.0.0.0 with credentialed CORS. Attackers can craft a malicious DNS rebinding page to issue authenticated requests to the local API server, reach the shell execution endpoint with a bash-enabled preset, and achieve remote code execution as the API process user while also overwriting LLM and data-source settings to exfiltrate credentials.
OpenHarness /issue and /pr_comments slash commands lack remote_invocable=False protection, allowing remote channel senders to write attacker-controlled Markdown into project context files. Admitted remote attackers can inject malicious content into .openharness/issue.md and .openharness/pr_comments.md files, which are subsequently injected into runtime system prompts, persistently influencing local agent behavior.
OpenHarness ohmo gateway /resume and /summary slash commands default remote_invocable to True, allowing admitted remote senders to enumerate and load arbitrary session snapshots by ID. Attackers can exploit this to access victim snapshots containing private prompts, credentials, tool output, and file paths via shared gateway channels.
nanobot is a personal AI assistant. In versions 0.1.5.post3 and prior, the WhatsApp bridge in bridge/src/whatsapp.ts constructs a filesystem path using the fileName field from an incoming WhatsApp document message without sanitization. The WhatsApp bridge downloads media attachments and writes them to disk using a filename derived from the sender's message via documentMessage.fileName, which is concatenated with a prefix and its raw value is passed directly to path.join(mediaDir, outFilename). Node.js path.join resolves .. components, allowing an attacker to escape the intended media/ directory by sending a document with a crafted fileName such as ../../../.ssh/authorized_keys. Because the attacker also controls the file content (the downloaded buffer), this is a write-anywhere primitive — both path and content are attacker-controlled. A fix for this issue is planned for version 0.1.5.post4.
CVE-2026-12203 is an information disclosure vulnerability in the HKUDS AI-Trader product affecting the Research Export component. The issue involves the /api/research/agents.csv endpoint, which could be manipulated remotely to disclose information without requiring authentication. The vendor has released a patch that restricts access to authenticated agents with the research_exports capability to mitigate this vulnerability. The product uses a rolling release model, so specific affected versions are not provided.
Nanobot prior to version 0.2.1 contains a denial of service vulnerability in the Matrix channel media download handler that allows authenticated room members to exhaust process memory and bandwidth by sending media events with missing or invalid size metadata. Attackers can send multiple concurrent Matrix media events with omitted or invalid declared sizes to trigger simultaneous large media downloads that fully materialize response bodies before post-download rejection, consuming process resources until service degradation occurs.
Nanobot prior to version 0.2.1 contains a server-side request forgery vulnerability in the Microsoft Teams channel handler that allows remote attackers to exfiltrate Bot Framework bearer tokens by supplying a forged activity with an attacker-controlled serviceUrl value. Attackers can poison the stored conversation reference by sending a crafted inbound activity to the Teams webhook, causing subsequent bot replies to transmit token-bearing Authorization header requests to an attacker-controlled host.
Nanobot prior to version 0.2.1 contains a server-side request forgery vulnerability in the web_fetch tool that allows remote attackers to reach internal or private network hosts by supplying a URL that redirects to a loopback or private address via a 3xx Location header. Attackers can exploit the automatic HTTP redirect following behavior in the httpx library to bypass initial URL validation and cause the runtime to send outbound requests to internal hosts before final resolved URL validation is applied.
DeepCode through commit c991dc2 contains a path traversal vulnerability in the SPA catch-all route in new_ui/backend/main.py that allows unauthenticated attackers to read arbitrary files by supplying percent-encoded path segments to the GET /{full_path:path} endpoint. Attackers can bypass Starlette's path normalization by encoding slashes as %2F and dots as %2E%2E, causing the joined path to traverse outside FRONTEND_DIST and exposing sensitive files such as SSH private keys, TLS certificates, and application secrets with a single HTTP request.
HKUDS OpenHarness contains a remote code execution vulnerability in the /bridge slash command that allows remote senders accepted by configuration to execute arbitrary operating system commands. Attackers can invoke the /bridge spawn command with attacker-controlled command text that is forwarded to the bridge session manager and executed through the shared shell subprocess helper, allowing them to spawn shell sessions as the OpenHarness process user and access local files, credentials, workspace state, and repository contents.
HKUDS OpenHarness prior to PR #147 remediation contains an insecure default configuration vulnerability where remote channels inherit allow_from = ["*"] permitting arbitrary remote senders to pass admission checks. Attackers who can reach the configured channel can bypass access controls and reach host-backed agent runtimes, potentially leading to unauthorized file disclosure and read access through default-enabled read-only tools.
HKUDS OpenHarness prior to PR #156 remediation exposes plugin lifecycle commands including /plugin install, /plugin enable, /plugin disable, and /reload-plugins to remote senders by default. Attackers who gain access through the channel layer can remotely manage plugin trust and activation state, enabling unauthorized plugin installation and activation on the system.