Bypasses / Earlier patch
Earlier patch
49 cases that got past this protection.
- 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.
- vLLM: Responses API bypasses trust_request_chat_template for chat_template_kwargs The Responses endpoint accepts request-supplied Jinja chat templates that skip the trust-flag check, exhausting CPU and memory.
- 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.
- vm2: Default VM can mutate host TypedArray and ArrayBuffer intrinsics after the host-prototype pollution fix A prior fix blocked sandbox writes to classic host prototypes but omitted typed-array and ArrayBuffer intrinsics, so sandbox code can still mutate host objects.
- vm2: Host-returned Promise rejection can bypass vm2's unhandled-rejection hardening and terminate the host process Sandboxed code can trigger an unhandled host Promise rejection that the bridge leaves unhandled, crashing the Node.js host process.
- vm2: util.getCallSites() bypasses GHSA-v27g-jcqj-v8rw host-frame redaction, leaks host call stack Sandboxed code reaches the host util module's getCallSites, exposing host file paths and stack frames that an earlier redaction fix was meant to hide.
- vm2 sandbox escape to host RCE via revisited host-wrapped AggregateError bypassing Error sanitization cycle short-circuit A host-wrapped AggregateError revisited during one traversal escapes error sanitisation, so sandboxed code receives a live host object and runs commands on the host.
- vm2: GHSA-m283-3h24-438v fix bypass leads to host RCE via call/apply indirection Calling Promise rejection handlers through call or apply indirection skips the sanitiser, letting a host error carrying a host object reach the sandbox and run commands.
- MCP Atlassian: Incomplete path traversal fix allows intra-CWD module overwrite and RCE (bypass of GHSA-xjgw-4wvw-rgm4) An incomplete traversal fix still lets attachment downloads overwrite application source modules in the working directory, so injected code runs on restart.
- MCP Atlassian: Path traversal in upload_attachment allows arbitrary file read (incomplete fix for CVE-2026-27825) The upload MCP tools open any caller-supplied file path without the confinement check that the download path received in an earlier fix.
- MCP Atlassian: Path traversal in upload_attachment allows arbitrary file read and exfiltration via MCP tool call An authenticated MCP client passes an arbitrary local path to upload_attachment, which reads the file and sends its contents to Atlassian, because the upload path skips the safe-path check.
- MCP Atlassian: Path Traversal / Arbitrary File Read in confluence_upload_attachment MCP tool (incomplete fix of GHSA-xjgw-4wvw-rgm4) An MCP upload tool opens any local file path a caller supplies and posts its contents to an attacker-chosen Confluence host, since the earlier fix only covered downloads.
- NLTK: Entity-expansion DoS (billion laughs) via remaining raw ElementTree parses Entity-declaring XML handed to several NLTK parsers expands exponentially in memory, exhausting resources, because those call sites bypassed the safe XML parsing helper.
- NLTK: Default ENFORCE=False Disables All pathsec Security Controls The security checks added for earlier CVEs only log warnings by default, so path traversal, unsafe URLs and unsafe pickle loading still proceed.
- PraisonAI workflow include bypasses tools.py autoload opt-in and executes included recipe code Including a local recipe in a workflow runs its tools.py module code even when the autoload opt-in is off, getting around earlier hardening.
- 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 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.
- Keras model loading is vulnerable to denial of service through HDF5 shape bombs A specially crafted .keras archive declares oversized HDF5 datasets, so loading it allocates unbounded memory and the process is killed.
- Flowise: CVE-2025-8943 Patch Bypass: npm_config_yes bypasses MCP environment variable blocklist (Unauthenticated RCE) Setting npm_config_yes through the MCP server environment makes npx auto-install and run an arbitrary package, because only four environment names are blocked.
- Open WebUI: Model meta.knowledge read-only file access can be upgraded to file write/delete A model's knowledge-file metadata lets a read-only grant satisfy later write and delete checks, because an earlier fix validated only read access.
- Keras: HDF5 virtual datasets can disclose local files A malicious .keras or .h5 model file with virtual datasets pointing at local HDF5 files makes Keras read those files when the model is loaded, because the earlier fix missed this check.
- BerriAI litellm: UI User Enumeration leads to System-Wide Information Disclosure A user listing endpoint returns all accounts without proper authorization, and an earlier incomplete fix left it exposed.
- PraisonAI: AgentOS remains unauthenticated after incomplete fix version and allows remote agent invocation Agent API routes accept requests without authentication, so any caller can invoke deployed agents, including tool-enabled ones, through the chat endpoint.
- Open WebUI: Stored XSS to Account Takeover via Model Profile Images Model profile image uploads accept SVG data URIs that earlier fixes missed, so embedded scripts run when the image URL is opened.
- vLLM: incomplete CVE-2026-22778 fix leaks PIL repr addresses via Anthropic router Malformed image bytes make the Anthropic-compatible router echo an unsanitised exception containing a heap address, because a later endpoint skipped the earlier fix.
- PraisonAI: Arbitrary code execution via unguarded `spec.loader.exec_module` in `agents_generator.py` - sibling of CVE-2026-44334 Module paths read from YAML configuration are executed through importlib at call sites missed by the earlier environment gate, running attacker-supplied Python code.
- PraisonAI vulnerable to unauthenticated arbitrary file read via MCP workflow.show, workflow.validate, deploy.validate An unauthenticated MCP tools/call to workflow.show passes a filename straight to file reads, since the containment fix covered other handlers but not this one.
- vm2 has a CVE-2023-37903 patch bypass: nesting:true without explicit require still allows full RCE Omitting the require option makes the nesting guard's strict comparison fail, so an inner sandbox gets the disallowed configuration and can spawn processes.
- Open WebUI has stored XSS via unsanitized Office/Excel/DOCX file preview rendering ({@html} without DOMPurify) Office file previews are rendered as raw HTML without sanitisation, a regression of an earlier fix, so uploaded documents can run script.
- n8n Has an XML Node Prototype Pollution Patch Bypass A prior fix for prototype pollution in the XML node can be circumvented by a workflow author, and combined with other nodes it reaches code execution on the host.
- Dockerfile command injection via envs[*].name in bentofile.yaml (sibling fix-bypass of CVE-2026-33744 and CVE-2026-35043) A malicious bentofile env name with embedded newlines produces unquoted Dockerfile RUN lines that execute during container builds, bypassing earlier quoting fixes.
- vm2 Access to Host Object Enables Sandbox Escape An incomplete earlier fix still exposes the host Object through symbols, letting sandboxed code reach the host realm and run child processes.
- PraisonAI has unauthenticated RCE via `tool_override.py` (CVE-2026-40287 patch bypass) The recipe runner dynamically executes a tools.py found beside a recipe template, and this loader path lacks the opt-in gate added to other tool loaders.
- VM2 Has Sandbox Breakout Through Promise Species A previous sandbox fix can be undone by overwriting built-ins used in its reset logic, letting sandboxed code reach the host process.
- n8n has SQL Injection in Snowflake and MySQL Nodes Table, column and update-key identifiers are interpolated into SQL in the Snowflake and legacy MySQL nodes, a gap the earlier fix did not cover.
- FastChat has Denial of Service Through Blocking Event Loop in Model Workers (Incomplete Fix for ff66426) Long-running generation requests in model worker APIs block the event loop, and an earlier fix missed several endpoints, so other routes still stall service.
- RAGAS has SSRF via Multi-Modal Faithfulness Collections Module The evaluator fetches URLs found in retrieved contexts without the SSRF guard added elsewhere, so attacker-supplied context can make the server send requests.
- PraisonAI has an incomplete fix for CVE-2026-34935 - OS Command Injection MCP server command parsing hands arbitrary executables and inline-code flags to subprocesses because the earlier fix added no allowlist or argument checks.
- PraisonAI: SQL Injection via unvalidated `table_prefix` in 9 conversation store backends (incomplete fix for CVE-2026-40315) The table prefix is interpolated into SQL in nine store backends that lack the validation added to the SQLite store, so crafted prefixes inject queries.
- LangSmith Client SDKs has Prototype Pollution in langsmith-sdk via Incomplete `__proto__` Guard in Internal lodash `set()` An incomplete __proto__ guard in a vendored lodash set lets keys using constructor.prototype pollute Object.prototype in the Node.js process.
- BentoML: Command Injection in cloud deployment setup script Unquoted system_packages values from bentofile.yaml are placed in a setup script run on cloud build infrastructure, a path an earlier quoting fix missed.
- vLLM introduced enhanced protection for CVE-2025-62164 Crafted sparse prompt embeddings reach PyTorch without index validation, so enabling the feature lets a user crash or stall the serving process.
- Local File Inclusion in mlflow Directory traversal sequences placed in a URI fragment evade the validation added earlier for query strings, letting requests read arbitrary local files.
- Code execution in pandasai Crafted dataframe descriptions lead the LLM to generate Python that the synthetic data executor runs, getting around earlier restrictions on code execution.
- langchain_experimental vulnerable to arbitrary code execution via PALChain in the python exec method Model-generated Python run through PALChain's exec can import modules via __import__, which the earlier filter failed to block.
- mlflow Path Traversal vulnerability An incomplete fix for an earlier MLflow path traversal flaw leaves a bypass that reaches files outside the intended directory.
- Code injection in `saved_model_cli` in TensorFlow The saved_model_cli input expression is passed to Python eval on an unsafe path, so supplied expressions run arbitrary code; the earlier fix missed this branch.
- Null pointer dereference in TFLite's `Reshape` operator A reshape input with no data buffer, supplied as a 1-D shape tensor, slips past an earlier null check and is dereferenced.