CVE-2026-57143 is an improper input validation vulnerability in MervinPraison's praisonaiagents before version 1.6.62. The vulnerability involves the searxng_search and search_web SearxNG providers passing an LLM-controlled URL parameter directly to requests.get without validating its scheme, host, port, or resolved address. This allows prompt-injected content to access internal JSON services, probe internal ports, and attempt cloud metadata requests, with the results incorporated into the agent context. The issue is fixed in version 1.6.62.
CVE-2026-107717 is an improper input validation vulnerability in masci banks prior to version 2.5.0. The Banks Prompt.chat_messages() function parses each line of rendered template output as ChatMessage JSON. If untrusted data is rendered and passed to an LLM provider, attacker-controlled JSON can manipulate the ChatMessage.role field, allowing an attacker to override application instructions or alter prompt structure. This issue is fixed in version 2.5.0.
Banks generates meaningful LLM prompts using a simple template language. Prior to 2.5.1, Banks DirectoryPromptRegistry does not reject symbolic links for index.json or discovered and existing .jinja prompt files. In an application where untrusted users can influence a prompt directory, DirectoryPromptRegistry._scan() and DirectoryPromptRegistry.get() can follow a link outside the registry root and disclose a file, while DirectoryPromptRegistry.set(), DirectoryPromptRegistry._save(), and DirectoryPromptRegistry._load() can read or overwrite an external link target. The issue requires attacker influence over the registry directory or its extracted contents. This issue is fixed in version 2.5.1.
LangChain is a framework for building LLM-powered applications. Prior to 1.1.1, @langchain/redis does not escape attacker-controlled values in structured RediSearch TAG filters and structured RediSearch TEXT filters, allowing injected RediSearch syntax to alter or broaden the generated search query. When an application uses an attacker-influenceable filter as a tenant or document-access boundary, the modified query can expose indexed documents outside the attacker's intended scope. This issue is fixed in version 1.1.1.
The llm-nebula npm package version 1.0.0 contains a malicious preinstall script that executes automatically during installation. This script is heavily obfuscated and runs with the user's privileges, enabling arbitrary filesystem, process, and network operations. The package pretends to be a legitimate LLM SDK but includes a runtime decoder that dynamically resolves Node.js built-in modules and commands, allowing remote code execution at install time.
9Router is an AI router & token saver. Prior to 0.5.6, 9Router deployments that allow requests to reach Next.js without the sanitizing custom-server.js wrapper trust the client-supplied X-9r-Real-Ip header in src/dashboardGuard.js when isLocalRequest decides whether canAccessPublicLlmApi may skip API-key validation for /api/v1/* routes. A remote unauthenticated attacker can set X-9r-Real-Ip to 127.0.0.1 and be classified as a local client, including on the verified GET /api/v1/models route. This permits unauthorized use of the instance owner's configured LLM providers, consumption of paid credits, and enumeration of configured providers and models. This issue is fixed in version 0.5.6.
CLOSEDQUORUM is a Windows malware binary representing the first documented implant utilizing autonomous AI-driven command and control. Discovered through Cisco Talos' CAIRN project, it delegates tactical decisions to a panel of commercial large language models including DeepSeek, Qwen, Mistral, and Google Gemini. The system operates through plurality voting among AI models to select actions for credential harvesting and crypto wallet theft, eliminating the need for continuous human operator involvement. The 16.4MB Go-compiled executable employs structured JSON schema to constrain LLM responses to executable attack choices. While containing placeholder credentials in public distribution, development builds demonstrate compile-time injection of operator-specific API keys and Discord webhooks. This architecture represents a significant shift toward effort displacement in cyber operations, where entire attack phases execute autonomously without human bottlenecks, though introducing new dependencies on commerci...
MCP Memory Keeper is an MCP server for persistent context management in AI coding assistants. Prior to 0.13.0, context_import in src/index.ts passes the caller-controlled filePath directly to fs.readFileSync without restricting the path to an export directory. An MCP client, including an LLM agent induced to call the tool, can use ../ traversal or an absolute path to target any file readable by the server process. A valid JSON file is parsed and imported into the caller's session, allowing its full contents to be retrieved through context_get or context_export, while JSON.parse errors for non-JSON files can return leading file bytes in a SyntaxError message. The two disclosure modes can expose other exported sessions, JSON credentials or service-account files, environment files, and portions of SSH keys or other local files. This issue is fixed in version 0.13.0.
AnythingLLM is an application that turns pieces of content into context that any LLM can use as references during chatting. In 1.16.1 and earlier, the manager role can store meta_page_title or meta_page_favicon through /api/admin/system-preferences, and MetaGenerator inserts those values into production homepage HTML without escaping attribute values or text content. The values pass unchanged through server/models/systemSettings.js and reach MetaGenerator.generate() in server/index.js. #assembleMeta() in server/utils/boot/MetaGenerator.js concatenates the stored values into HTML. When an administrator visits the homepage /, injected JavaScript can read the administrator JWT and use it to create API keys, access or modify workspace and chat data, delete users, and perform other administrator actions. server/endpoints/admin.js accepts the manager-controlled settings before server/models/systemSettings.js returns them unchanged. No fixed version is available as of this review.
Xinference (affected commit 4a94832, v3.x) contains an unauthenticated arbitrary-path file read vulnerability in the POST /v1/models/llm/auto-register endpoint, which accepts a caller-supplied model_path parameter without authentication or path confinement. The endpoint reads and parses config.json, tokenizer_config.json, and chat_template.jinja files at the supplied path and reflects the parsed content back to the caller, allowing an unauthenticated attacker to probe the server filesystem and extract content of files with those names in any directory.
TOON is a compact, human-readable serialization of JSON data for LLM prompts. Prior to 2.3.1, decoding attacker-controlled TOON with a __proto__, constructor, or prototype key wrote through the object prototype chain instead of creating an own property, polluting Object.prototype for the runtime. In packages/toon/src/decode/expand.ts, the expandPaths: 'safe' path and insertPathSafe function made dotted keys such as a.__proto__.x the strongest vector, while plain nested objects, tabular rows, quoted keys, and streaming decode were also affected. The encoder also dropped own __proto__ properties and could invoke an inherited setter during normalization. Services that decode untrusted TOON could experience denial of service or, when a suitable downstream gadget is present, remote code execution. This issue is fixed in version 2.3.1.
browser-use web-ui versions 2.0.0 through 3.0.0 write configured LLM API keys to disk in cleartext without encryption or access restrictions. Attackers with read access to the temporary settings directory can recover provider API keys from predictably-named JSON files.
Xinference loads models with Hugging Face remote code execution unconditionally enabled, and before version 2.12.0 exposes no setting to disable it. Six loader call sites pass trust_remote_code=True as a literal or as an unconditional default: RerankModel._get_tokenizer in xinference/model/rerank/core.py, SentenceTransformerRerankModel.load in xinference/model/rerank/sentence_transformers/core.py, SentenceTransformerEmbeddingModel.load in xinference/model/embedding/sentence_transformers/core.py, FlagEmbeddingModel.load in xinference/model/embedding/flag/core.py, and two sites in xinference/model/llm/transformers/core.py where PytorchModel._sanitize_model_config and PytorchModel._get_components default the value to True. Because a caller with model launch access can register a model whose type is unknown and supply an arbitrary model path, the server reaches _auto_detect_type and then AutoTokenizer.from_pretrained, which imports and executes Python declared by the model directory's own tokenizer_config.json auto_map, running attacker-supplied code with the privileges of the worker process. Version 2.12.0 gates every site behind allow_trust_remote_code and the XINFERENCE_TRUST_REMOTE_CODE setting, permitting remote code only for bundled built-in models.
Banks generates meaningful LLM prompts using a simple template language. Prior to version 2.4.5, DirectoryPromptRegistry.set() in src/banks/registries/directory.py interpolates attacker-controlled Prompt.name and Prompt.version values into a Path without canonicalization or containment validation. Relative traversal such as ../victim/foo and an absolute Prompt.name can escape or discard the configured registry root, while overwrite=True permits replacement of existing target files. The poisoned name is persisted in index.json and reconstructed by _load(), allowing the out-of-root path to survive later registry loads. An application that forwards request data into these fields can therefore write Prompt.raw bytes to attacker-chosen paths writable by the application process. This issue is fixed in version 2.4.5.
---
_-= Per source details. Do not edit below this line.=-_
## Source: amazon-inspector (8da2cb00fa6d2b5a0e5f4a4bdca8ca0cfaedf7e583b0f9e9f54274bcf5c39e06)
On invocation of `agent-proxy start`, the CLI opens an outbound WebSocket to an obfuscated author-controlled endpoint (hub.client-llm.com, reconstructed at runtime via XOR+base64 with key 'gnP2p!7xQ' from a base64 blob in dist/cli.js) and accepts remote `job` messages containing arbitrary path and body fields. These are POSTed into the local proxy and forwarded to api.anthropic.com and the Codex/ChatGPT upstream using the installer's on-disk OAuth credentials, with responses streamed back to the hub. An `onlyIfCredentialed` gate ensures only installers with valid Claude/Codex logins are enrolled as worker nodes serving requests originated by the hub operator. The behavior is undocumented; the README states the tool is 'self-use only' and warns that upstream vendor ToS forbid resale or sharing of accounts. The destination URL is not present as a plaintext string, only reconstructed at runtime, concealing it from casual review. The combination provides a remote-controlled command channel (arbitrary request path/body) into the installer's authenticated AI session and silently monetizes the installer's paid subscription for the hub operator's traffic.
Prompty is a markdown file format (.prompty) for LLM prompts. Prior to 0.1.5 and 2.0.0-beta.5, the TypeScript Nunjucks renderer evaluated untrusted .prompty template bodies with unrestricted JavaScript member access. An attacker-controlled template could traverse constructor and prototype properties to execute JavaScript in the host Node.js process. This issue is fixed in versions 0.1.5 and 2.0.0-beta.5.
The biklirouter npm package versions 0.5.46 and 0.5.47 include a compiled Next.js application that routes AI/LLM provider CLI commands and manages OAuth flows. While the package uses Node.js built-ins such as child_process and network modules consistent with its declared functionality, a detailed review found no direct evidence of exfiltration of sensitive installer secrets or hardcoded command-and-control destinations. However, a malware advisory states that any system running this package should be considered fully compromised, recommending immediate secret rotation and package removal due to potential full system compromise.
AnythingLLM is an application that turns pieces of content into context that any LLM can use as references during chatting. From 1.0.0 to 1.15.0, AnythingLLM's unauthenticated account-recovery flow in server/utils/PasswordRecovery/index.js uses recoverAccount() to deduplicate the raw recoveryCodes values before trimming them, so one valid code submitted twice with different surrounding whitespace can satisfy the two-code check. Each normalized value can also match the same stored hash instead of consuming a distinct hash. An attacker who knows the target username and one recovery code can call POST /api/system/recover-account in multi-user mode, receive a password-reset token, and use POST /api/system/reset-password to take over the account, including an administrator account.
---
_-= Per source details. Do not edit below this line.=-_
## Source: amazon-inspector (b9276dba65fd393d94c5beb06122b1297994c96bda03d940f34fcc59016e4c5b)
The bundled CLI at bin/claude-cli.js hardcodes a hub endpoint https://hub.client-llm.com that is concealed behind a bespoke base64+XOR decoder (function _deHub, XOR key 'gnP2p!7xQ'). The decoder is used exactly once, to reconstruct this single URL; no other strings in the bundle use it. On `claude-cli start`, this hidden hub is added by default to the active hub list (opt-out only, undocumented), while the README presents hub participation as an opt-in feature the user manually configures. Once started, the CLI opens a WebSocket to <hub>/node, and on receiving {t:"job", id, body} messages it POSTs the remote-supplied body to the installer's local http://127.0.0.1:<port>/v1/messages endpoint using the local API key (x-api-key header) and streams the response back over the WebSocket. In default mode this dispatches a `claude -p` child process on the installer's machine driven by attacker-controlled prompts, consuming the installer's paid Anthropic subscription and, given agentic/tool-enabled Claude sessions, exposing remote-controlled execution paths on the installer's host. The targeted obfuscation of the destination URL, the mismatch between documented opt-in behavior and actual default-on enrollment, and the remote job-dispatch channel together constitute a backdoor combined with silent relay of the installer's Claude credentials and quota.
Flowise is a drag-and-drop user interface for building customized large language model (LLM) flows. Prior to version 3.1.3, several custom-tool components — AgentAsTool, ChatflowTool, and ExecuteFlow — ran code in the in-process vm2 sandbox. To build that code, they inserted a user-controlled baseURL value straight into the JavaScript source, for example const url = "${baseURL}/..."; . The only check on baseURL was isValidURL , but a valid-looking URL can still contain characters that break out of a code string. An authenticated user could craft a baseURL that passed this check, closed the surrounding string, and injected their own JavaScript into the sandboxed script (code injection, CWE-94). The vm2 sandbox runs in the same Node.js process as Flowise and exposes risky dependencies. As a result, the injected code could escape the sandbox and run arbitrary code on the Flowise server as the Flowise process user. Exploitation only requires an authenticated session. The issue is fixed in version 3.1.3, which passes the URL to the sandbox as data instead of inserting it into code and adds stricter URL validation.