CVE-2024-34359
No source can show you its own past. This one can.
Print this page to keep the view as it stands on this date. Everything hidden behind a disclosure opens, and every receipt prints with its value.
llama-cpp-python vulnerable to Remote Code Execution by Server-Side Template Injection in Model Metadata
rank 32 = 7 + 24 + 1 + 0 + 0
rank@v1 · scheduled
Someone is already using it
CISA KEV catalog · public proof-of-concept · FIRST EPSS
EPSS 0.26001 · without this subject the record ranks 25
What it costs you if it fires
NVD CVSS base score, then the CNA's own
CVSS 9.7, from the CNA · without this subject the record ranks 8
It is getting worse, not staying still
FIRST EPSS series, rescorings and bad loads removed
rose on 2026-04-14 · without this subject the record ranks 31
How much of the world it touches
NVD CPE configurations
0 products in the CPE ranges · without this subject the record ranks 32
Someone else already set the date
CISA KEV remediation due date
no CISA due date · without this subject the record ranks 32
$1.30M = $4.99M incident cost x 1 exposure x 0.26001 likelihood
udv@v1 · unaudited · wrong for you until you say what you run
$4.99M is the cost of a breach that happened, from IBM Cost of a Data Breach Report 2026, global average. It is not the value of any asset of yours. That average already covers breaches of every severity, so scaling it by exposure understates a partial-impact record. That is an assumption, not a measurement. Your own asset value goes in the decision record.
Hindsight · the state
Foresight · where to check next
No weakness class or product neighbourhood is held for this record, so this index cannot name anywhere to check next. That is an absence in the record, not a finding about the software.
Running abetlen llama-cpp-python? Check your version
A deterministic range check against the ranges this record publishes. It tells you where a version sits in the record, not whether your deployment is exploitable. No account needed.
Watch this record
Matches your own product list against every record this index holds, and tells you when one of them moves.
Set up a watchlist (needs a Practitioner account)
Record your call
Keep your verdict, your reason and the evidence together. The record takes its date when you save it.
AcceptDeferPatch by…Not affectedMitigatedChoose your verdict, then sign in or use an email link. Your first 100 records are free. Evidence is frozen when the record is created.
Private to your organization unless you share it. Put the proof link in your ticket so a reviewer can inspect the basis for your call. See how decision records work.
Harm, if it fires from the vector and the weakness
Written for the ticket you're about to file.
CISA KEV & Federal Remediation Deadline from the record and CISA
The fix, and how long it's been waiting.
Exploit Status & Known PoCs a ladder, not a score
Current rung: Nothing observed in held sources.
- RUNG 1Nothing observed in held sources
- RUNG 2Proof-of-concept published
- RUNG 3Exploited in the wild · CISA KEV listed
- RUNG 4Ransomware campaign use known
Affected & Fixed Versions as published
A rung lights from a row in the index, never from an assessment. "Held sources" is the true scope of the claim: this index reads the CVE record, NVD, CISA KEV, FIRST's EPSS and public proof-of-concept repositories, and it does not observe attacks.
Attack path
What it is stated by the CNA
llama-cpp-python is the Python bindings for llama.cpp. `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. receipt
Read the rest of the CNA's description
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 remote code execution by a carefully constructed payload.
NVD's analysis a second authority
- weakness
- CWE-76
- also scored by
- [email protected] 9.6 CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:C/C:H/I:H/A:H
check it yourself github.com github.com receipt
Dated events
Movement the material steps
This score more than doubled inside 72 hours. Both observations are in the full history below.
Show all 32 observations
32 observations held. A highlighted step crossed a promotion threshold. A step marked with a scoring model moved because the instrument was replaced, not because the vulnerability changed. receipt
The full biography
Exploit probability more than doubled between 2025-11-18 and 2025-11-21 (0.01288 → 0.5917) and the record had been public for 2 years, 4 months when this index was last written.
Siblings and hubs
Records related to this one
Browse the index from here
These are index-wide views, not findings about CVE-2024-34359.
What usually comes next
Everything above needs no account. What is below is work: assembling it, scoping it to your estate, watching it, and handing it to someone else.