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: Path Traversal in save_maxent_params / load_maxent_params Allow Arbitrary File READ/WRITE The maxent save and load functions open files under a caller-supplied directory without the library's path restriction check, allowing reads and writes elsewhere.
- NLTK dot2img runs Graphviz 'dot' without the trusted-exec chokepoint, enabling local code execution via a writable PATH element A partial earlier fix left one Graphviz call resolving dot from PATH without trust checks, so a planted binary in a writable directory runs.
- NLTK: Symlink escape in chat80.sql_query bypasses the pathsec data-root sandbox to read out-of-root SQLite databases Symlinks inside a trusted data root are followed by sqlite3 because lookup never resolves real paths, letting queries reach databases outside the sandboxed roots.
- NLTK: Quadratic-time CPU denial of service in CCG lexicon parser (fromstring) on wide bracket bodies and operator chains A lexicon parser repeatedly copies remaining input while scanning bracket bodies and operators, so long flat inputs cause quadratic CPU consumption.
- NLTK: Quadratic-time CPU denial of service in Tree.chomsky_normal_form right-factor binarization via list.pop(0) A wide flat parse tree makes right-factor binarization quadratic in child count, so a single crafted tree from untrusted input stalls the CPU.
- NLTK: Quadratic-time CPU denial of service in ConllCorpusReader SRL instance builder via nested span rescan Parsing a crafted CoNLL SRL block forces quadratic-time span rescans for each predicate, consuming large amounts of single-threaded CPU.
- NLTK: Unbounded CPU consumption in SteppingRecursiveDescentParser.parse() with ambiguous or self-embedding grammars The stepping recursive descent parser runs without a time or step limit, so an ambiguous grammar with short input can pin a CPU core indefinitely.
- NLTK: Quadratic CPU exhaustion in TreebankWordTokenizer.span_tokenize via list.pop(0) in quote-restore loop Repeated list removal inside the quote-restoration step lets a long run of quote characters consume a CPU core for a prolonged period.
- NLTK: BllipParser passes caller-supplied model paths to the bllipparser C extension with no pathsec check, enabling arbitrary read BllipParser passes caller-chosen paths straight to a native extension, skipping NLTK's path validation, so arbitrary readable files can be opened.
- NLTK: Quadratic CPU exhaustion in LegalitySyllableTokenizer.tokenize via unguarded onset reversal loop Passing one long all-vowel token to the syllable tokenizer triggers a quadratic loop that keeps a CPU core busy for the duration of the call.
- NLTK: Symlink escape in YCOECorpusReader lets a psd/pos directory symlink read files outside the corpus root within the data sandbox A directory symlink inside a corpus lets nested readers escape the corpus root and read files elsewhere in the data sandbox.
- Kubeflow Pipelines: Cross-tenant execution-cache poisoning / output substitution via a namespace-blind cache key in the cache webhook (multi-user) The pipeline step cache key omits the tenant namespace, letting one tenant read or pre-populate cached outputs belonging to another tenant.
- Kubeflow Pipelines — API server fetches a caller-supplied pipelineURL without host validation The pipeline API server fetches a caller-supplied package URL without restricting destinations, letting authorised users make requests from its network position.
- Kubeflow Pipelines: Reflected XSS in the frontend pod information and events error responses Pod name and namespace query values are echoed unescaped into HTML error responses, so a crafted link runs script in the frontend origin.
- Kubeflow Pipelines: RBAC bypass on shared pipelines allows any authenticated user to delete or modify global pipelines Multi-user mode skips authorisation for shared pipelines with an empty namespace on write and delete, letting any user modify or remove them.
- Kubeflow Pipelines: SQL injection through unvalidated pagination tokens in Kubeflow Pipelines API server Crafted base64 pagination tokens inject SQL through unvalidated identifier fields in list API calls, letting attackers read database rows across namespaces.
- Kubeflow Pipelines API server lacks authorization for shared pipeline uploads Upload endpoints skip the access check when the namespace is empty, letting callers create or replace shared pipeline versions.
- Kubeflow Pipelines: Pipeline archive upload allows decompression bomb to exhaust server memory Pipeline archive uploads are decompressed without bounding the expanded size, so a small zip or tar file can exhaust API server memory.
- Kubeflow Pipelines: Run-log endpoint permits unauthorized access to pod logs The run-log endpoint returns pod logs without authorising the caller against the run, and accepts arbitrary pod names, exposing logs from unrelated workloads.
- LangGraph: Unencoded identifiers in LangGraph SDK v3 streaming can redirect authenticated GET requests Unencoded thread_id or assistant_id values containing relative segments are normalised into other server paths, so an authenticated GET is sent to a different endpoint than intended.
- ONNX: Out-of-Bounds Read in RNN/GRU/LSTM Shape Inference (Old Opsets, Missing Rank Check) A malformed old-opset RNN node with a short input shape makes shape inference read past array bounds, crashing any process that loads the model.
- ONNX: Null Pointer Dereference in ScanInferenceFunction Due to Missing Required Attribute Shape inference for the Scan operator dereferences a missing required attribute and accepts negative counts, crashing when given a malformed model.
- vLLM: Untrusted EPS inputs can trigger denial of service or Ghostscript RCE Image bytes from a request are autodetected by Pillow without a format allowlist, so an EPS file labeled as PNG reaches Ghostscript and runs commands.
- vLLM: Permissive default CORS origins allow cross-origin requests to the vLLM HTTP server The server defaults to a wildcard CORS origin and has no auth unless an API key is set, so any web page can read responses from a reachable instance.
- LiteLLM: Privilege Escalation to Proxy Admin via Cross-Domain Reuse of the Salt Key Reusing one encryption key for both secret storage and session tokens lets an internal user forge an admin token via a crafted key metadata field.
- LiteLLM: internal_user_viewer arbitrary local file read via `vertex_ai_credentials` An authenticated viewer sends an alias of a credentials field that the request blocklist misses, so a local file is read and sent out.
- LiteLLM vulnerable to reflected XSS in /sso/debug/callback An identity provider returns a display name containing a closing script tag, which the unauthenticated SSO debug callback reflects into inline script and executes.
- MLflow vulnerable to remote code execution on unauthenticated servers via third-party scorer deserialization importing caller-placed artifact modules On an unauthenticated MLflow server, scorer deserialisation imports a request-chosen module from an uploaded artifact and instantiates its class, running attacker code.
- MLflow vulnerable to arbitrary file read on unauthenticated servers via prompt model version source Prompt model version sources without a URL scheme skip path validation, so artifact requests can read arbitrary server files, including the auth database.
- NLTK: WekaClassifier passes an unvalidated caller-controlled model path to the weka JVM, bypassing nltk.pathsec containment The WekaClassifier wrapper passes caller-supplied model paths to the Java tool unchecked, skipping the containment checks that other NLTK loaders apply.
- LangGraph: Namespace labels containing `.` reach other namespaces in Postgres and SQLite stores Namespace label validation runs only on writes, so dotted labels in reads, deletes and batches reach other namespaces' stored data.
- MLflow: SSRF on unauthenticated MLflow servers via third-party scorer kwargs in the scorer invoke, register, and GenAI evaluate endpoints An unauthenticated endpoint instantiates a class named in JSON whose module allowlist does not constrain the class or its arguments, forcing outbound requests to attacker-chosen hosts.
- MLflow FileStore trace IDs allow path traversal and arbitrary file writes A crafted trace ID with traversal sequences, sent unauthenticated to the trace endpoint, makes the FileStore backend write metadata files outside the experiment directory.
- PraisonAI: Call API localhost-only authentication bypass via spoofed Host header The localhost-only exemption from auth trusts the request Host header, so a spoofed Host value lets remote callers invoke agents unauthenticated.
- PraisonAI: Jobs API is unauthenticated by default and allows attacker-controlled webhook SSRF The jobs API is open unless a key variable is set, and the webhook URL is validated once then re-resolved at connection time, allowing rebinding to internal hosts.
- PraisonAI: API deploy code generator embeds unescaped YAML fields into Python source Host and file path values from a YAML config are pasted unescaped into generated Python source, so injected expressions run when the API server starts.
- PraisonAI: DNS rebinding bypass in `web_crawl` SSRF protection allows internal response disclosure The URL guard resolves hostnames only at check time while the fetcher resolves them again, so a rebinding domain returns internal content to the crawl tool.
- PraisonAI: Project config can auto-save agent output outside the project root A repository config can set an absolute or traversal output path, so agent responses are written to files outside the project root.
- Pydantic AI: SSRF cloud-metadata blocklist bypass via IPv6 zone identifiers The cloud-metadata blocklist compares IPv6 addresses without considering zone identifiers, so a zone-scoped metadata address passes and reaches the endpoint.
- Docling: Unbounded table rowspan/colspan in HTML, JATS, ODS and BoxNote backends causes CPU/memory exhaustion A tiny HTML or spreadsheet file with huge span attributes makes the converter loop and allocate oversized tables, exhausting CPU and memory.
- Docling: Crafted DoclingDocument JSON embeds local image files into converted output A crafted serialized document sets an image reference to a local path, and the converter reads that file and embeds it in exported output.
- Docling has SSRF guard bypass in remote resource fetching (DNS rebinding / multi-record resolution; no IP validation in HTML render mode) Private-address checks resolve hosts differently from the HTTP client, so DNS rebinding or backslash parsing reaches internal addresses.
- Docling has arbitrary file read/write (and command execution when shell-escape is enabled) when rendering untrusted TikZ with the opt-in Tectonic engine A crafted LaTeX/TikZ input rendered via Tectonic can read and write arbitrary files, and run shell commands when shell escape is on.
- Langflow : Cross-user flow access and vertex execution via deprecated /api/v1/build/{flow_id}/vertices endpoints Deprecated chat routes accept a flow ID without checking ownership, so any logged-in user can read and execute another user's flow vertices.
- PraisonAI: AgentMail webhook mode accepts forged unsigned message.received events and invokes agents A public webhook endpoint accepts unsigned message events and dispatches them to agents, so anyone can forge inbound messages that trigger agent replies.
- PraisonAI: Code Injection via f-string Interpolation in Deploy API Server Generation An unsanitised agents file value is interpolated into generated Python server code, which is then executed, so injected commands run.
- PraisonAI: AgentMail webhook lacks signature verification, allowing unauthenticated message injection and sender spoofing The AgentMail webhook accepts unsigned requests and trusts the sender field in the body, so anyone can inject messages as any address.
- Flowise NodeVM sandbox escape via puppeteer allowlist - authenticated RCE and arbitrary file read via Chromium An authenticated user passes executable path and args to puppeteer.launch from the sandboxed custom function, spawning host commands outside the sandbox.
- vm2 contains a sandbox escape vulnerability The NodeVM denylist omits child_process, so sandboxed code allowed to load builtins can require it and run host commands.
- Docling: METS-GBS archive member limit enforced after full member enumeration (memory exhaustion during format detection) Tar archives have all member headers enumerated into memory before the member-count limit applies, so many empty entries exhaust memory.