Browse / AI Framework and Platform Flaws
AI Framework and Platform Flaws
Classic software bugs in the libraries and platforms that run AI: memory corruption, traversal, SSRF, auth. 1,542 records. Reference: CWE.
- Denial of service 377
- Authentication and access control 256
- Path traversal and file access 192
- Memory corruption 159
- Code and command execution 133
- Server-side request forgery 103
- Information exposure 78
- Sandbox escape 76
- Cross-site scripting 68
- Other weaknesses 46
- SQL and query injection 22
- Real-world incidents 10
- Cross-site request forgery 8
- Unsafe deserialisation 8
- Data exfiltration 6
- NLTK: Uncontrolled search path when invoking the Graphviz 'dot' binary The dot binary is launched by bare name, so a planted file named dot in the working directory or a PATH entry runs instead of Graphviz.
- NLTK: JVM argument injection bypass via per-call options in the NLTK Stanford wrappers (incomplete fix of CVE-2026-12841) Per-call Java options skip the validator added for an earlier fix, letting callers pass agent flags that run code inside the JVM.
- LiteLLM vulnerable to server-side template injection in the /prompts/test endpoint The prompt test endpoint renders attacker-supplied template text with an unsandboxed Jinja2 environment, letting unauthenticated callers run OS commands.
- PraisonAI: [Auth Bypass] PraisonAI async Jobs API (`/api/v1/runs`) has no authentication — unauthenticated job execution, result theft, cancel and delete The jobs API router has no authentication, so any reachable caller can submit, read, cancel and delete agent jobs.
- PraisonAI serve agents --api-key is ignored, allowing unauthenticated remote agent execution The configured API key is never checked on the agent invocation endpoints, so unauthenticated network callers can run agents.
- PraisonAI: Authentication fail-open in Recipe server allows unauthenticated access when API key or JWT auth is configured without a secret The recipe server skips API key and JWT verification when no secret is configured, so requests are accepted without credentials even on non-local interfaces.
- PraisonAI vulnerable to Server-Side Request Forgery via DNS rebinding bypass in webhook_url validation The webhook URL check ignores DNS lookup failures and never rechecks the resolved address before the server posts to it, so a rebinding domain can reach internal hosts.
- PraisonAI: `--api-key` flag on `praisonai serve` is not properly enforced The serve command's --api-key option is parsed but never applied to the FastAPI app, so agent endpoints accept unauthenticated requests.
- PraisonAI: [Auth Bypass] `praisonai serve agents --api-key` is silently ignored — agent-invocation routes (`POST /agents`, `POST /agents/{agent_name}`) run unauthenticated Agent invocation routes never check the configured API key or call the existing token verifier, so unauthenticated callers can run registered agents.
- PraisonAI: Webhook SSRF via DNS fail-open in `JobSubmitRequest.validate_webhook_url()` — bypass of CVE-2026-40114 Webhook URL validation accepts hostnames that fail DNS lookup, and the later delivery resolves them again, so a record can be repointed to internal addresses.
- praisonaiagents: AgentServer declares auth_token but never enforces it on any route AgentServer stores an auth_token setting but no route checks it, so unauthenticated clients can reach every endpoint.
- praisonaiagents vulnerable to arbitrary file write via unsanitized `user_id` in `FileMemory.__init__()` — path traversal to any writable location The user_id parameter is joined into a base path without sanitisation, so traversal sequences let callers write memory files to arbitrary writable locations.
- praisonaiagents has an SSRF protection bypass in `spider_tools._host_is_blocked()` via DNS-resolved hostnames (`127.0.0.1.nip.io`) The spider tools blocklist checks only literal hostnames, so a name like 127.0.0.1.nip.io resolving to loopback reaches internal services.
- NLTK CrubadanCorpusReader path traversal allows arbitrary file disclosure A corpus mapping value that is an absolute path makes the reader open files outside the corpus, skipping the containment check.
- NLTK TweetTokenizer vulnerable to denial of service through catastrophic regex backtracking Crafted strings of repeated separators make a tokenizer regex backtrack exponentially, so a few kilobytes of text can stall a worker thread for minutes.
- New API: User List API Leaks Root User Access Token Leading to Privilege Escalation The admin user list API serialises each user's access token, so an admin can read the root account's bearer token and act as root.
- New API: Integer overflow in quota billing yields negative charges (self-crediting) Oversized quantity fields wrap during integer conversion in billing, producing a negative quota that credits the attacker's balance at settlement.
- New API: Admin can reset passkeys for same-level or higher-privileged users A lower-privileged admin can reset passkeys of same-level or higher accounts because the reset endpoint skips the role comparison used by sibling endpoints.
- New API: Unauthenticated payment webhooks allow memory and disk DoS via unbounded body reads and full-body logging Unauthenticated payment webhook routes read and log entire request bodies before signature checks, so oversized requests exhaust memory and disk.
- MLflow: Unauthenticated full-read SSRF in webhook delivery: _validate_webhook_url bypassed via unvalidated HTTP redirects (and DNS rebinding) Webhook delivery validates the initial URL but follows redirects without re-checking the target, letting an unauthenticated caller reach internal addresses.
- vm2: Sandbox Breakout Using Dangerous Host Proto Mutators Calling a blocked host mutator through an indirect call form evades a prior fix, letting sandboxed code reach host objects and run commands.
- vm2's bufferAllocLimit cap bypassed by Buffer.concat and Buffer.from arrayLike Sandboxed code can allocate unlimited host memory through Buffer.concat and array-like Buffer.from, skipping the configured allocation cap.
- VM2 has Missing Error.cause Sanitization that Enables Sandbox Escape to RCE Host error objects keep a cause property that the sandbox sanitizer never cleans, so sandboxed code can reach host objects like process and run commands.
- vm2: NodeVM `builtin: ['*']` exposes `os` and `dns` — process-wide observability reads AND writes that hijack the host (sibling class of GHSA-9g8x-92q2-p28f) Enabling all builtins in the NodeVM sandbox exposes the os and dns modules, letting sandboxed code read host details and change process-wide DNS servers.
- vm2 has Memory Exhaustion DoS via bufferAllocLimit Bypass Allocating memory through ArrayBuffer and typed arrays skips the Buffer size cap added by an earlier fix, letting sandboxed code exhaust host memory.
- nltk: Arbitrary File Read via Path Traversal in nltk.data.load() through Percent-Encoded Sequences Percent-encoded dot segments pass path checks that run before url2pathname decodes them, letting resource lookups escape the NLTK data directory.
- Pydantic AI: Unvalidated UploadedFile references in UI adapters allow server-side file access using the application's credentials Client-supplied file references in message history skip URL validation and reach the provider, which reads them with the server's own credentials.
- NLTK network URL validation permits SSRF to RFC 6598 shared-address-space hosts The private-address check misses the RFC 6598 shared range, so URLs pointing there pass validation and are fetched by the server.
- LangGraph: Namespace prefix matching crosses segment boundaries in Postgres and SQLite stores Postgres and SQLite stores match flattened namespaces with LIKE prefixes, so scoped reads return items from sibling namespaces such as foo and foobar.
- MLflow AI Gateway permits SSRF through an unvalidated api_base Gateway secrets accept an unvalidated base URL that the proxy later requests, letting any authenticated user reach internal addresses and metadata services.
- Open WebUI: A folder write-collaborator can permanently delete the owner's chats by deleting a shared subfolder Write access on a shared subfolder lets a collaborator delete the owner's chats, because the delete handler checks write access rather than ownership.
- Open WebUI: Tool source code disclosed to read-only users via the tool list and get endpoints Read endpoints for shared tools return the full Python source to non-writer users, exposing any credentials or internal URLs embedded in the code.
- Open WebUI: Any authenticated user can cancel another user's chat generation via the chat delete endpoint The chat delete endpoint cancels another user's in-flight generation tasks before checking whether the caller may delete that chat.
- Open WebUI: Same-origin XSS to account takeover via terminal file-preview iframe hardcoding allow-same-origin Terminal-served HTML previews run in an iframe that grants same-origin access, so scripts in the file run in the app origin and can read the session token.
- Open WebUI: Unapproved accounts can open terminal sessions via a WebSocket auth path missing the role check The terminal WebSocket authenticates its own token and skips the verified-user role check, so pending accounts can open shell sessions.
- Open WebUI: Instance-wide stall via automation recurrence rules that force multi-second parsing Minutely recurrence rules force the parser to enumerate decades of occurrences synchronously on the event loop, stalling the server for every user.
- Open WebUI: Any authenticated user can reach internal services and cloud metadata via NAT64-encoded URLs A NAT64-wrapped IPv4 literal passes a globally-routable check that only inspects the IPv6 form, letting the server fetch internal and metadata endpoints.
- Open WebUI: Users denied the image-generation permission can still generate images via chat completions Chat completion requests can include an image feature flag that the server honours without rechecking the user's revoked permission, so generation still runs.
- Open WebUI: DNS Rebinding SSRF Bypass URL validation resolves a hostname once, but the HTTP client resolves it again at connect time, so attacker DNS can rebind requests to internal addresses.
- Open WebUI: Deletion of directories and file embeddings in other knowledge bases via sync cleanup A user with write access to one knowledge base can delete directories and embeddings in other knowledge bases because object ids are never checked for ownership.
- Open WebUI: Any member with write access to a standard channel can edit or delete other members' messages On standard channels, the message update and delete handlers never checked authorship, so any member with write access could change or delete others' messages.
- Open WebUI: Account takeover via OAuth token exchange accepting tokens issued to any client The OAuth token exchange accepts provider access tokens issued to other client applications, so tokens from unrelated apps can mint sessions for victims.
- Open WebUI: SSRF into internal services via unvalidated sub-resource requests in the Playwright web loader Address validation runs only on the top-level page, so JavaScript in the loaded page can fetch internal addresses that are then returned to the user.
- Flowise: RCE via NodeVM Sandbox Escape in executeJavaScriptCode() nodeVMOptions Override Caller-supplied VM options spread over secure defaults re-enable blocked built-in modules, letting an authenticated user escape the sandbox and run system commands.
- Flowise: Authenticated arbitrary file write in the `S3 Directory` document loader via unsanitized S3 object keys Attacker-chosen S3 object keys are joined into local paths without stripping traversal sequences, so files are written outside the temporary directory.
- Flowise RCE via TypeORM DataSource Authenticated users can pass arbitrary options into a TypeORM DataSource through node settings, which can be abused to execute code on the server.
- Flowise Sandbox Escape to RCE A JavaScript escape from the Flowise custom function sandbox lets user code reach the host process and run arbitrary commands.
- Flowise RCE via SQLite Record Manager Node The SQLite record manager node lets additionalConfig override its database path, enabling attacker-chosen file writes that lead to code execution.
- Flowise: IDOR vulnerability exists at the GET /api/v1/organization/customer-default-source endpoint An authenticated user can request another customer's payment and profile data by changing the customerId parameter, because object-level ownership is not checked.
- Flowise: Unauthenticated Property Injection into Flow Execution Context via Ungated `overrideConfig` Spread in Prediction API The unauthenticated prediction endpoint spreads caller-supplied overrideConfig into flow context without the override gate, letting attackers inject session and chat-related properties.