Hugging Face

hybrid · United States , New York · founded 2016

Employees: 201-1000 · Stage: Series D+ · Funding: $395,000,000

24.1

moderate

default-balanced
composite score

Open-source AI platform and model hub that hosts over one million models, datasets, and spaces. Develops proprietary models (BigScience BLOOM collaboration, SmolLM, Zephyr) while serving as the primary distribution platform for the open-source AI ecosystem.

Analyst summary

Hugging Face is the de facto platform for open-weights models, datasets, and ML tooling. For enterprises, the key question is not Hugging Face itself but which models they host and run: the platform is a marketplace, not a single-model vendor. SOC 2 and GDPR posture is solid for the Hub and Enterprise services.

The platform of record for open-weights ML; the per-model risk assessment is still yours to do.

Rating: acceptable

Compliance posture

SOC 2 Type II	Yes (report 2025-04-01)
ISO 27001	No / not disclosed
ISO 42001 (AI management system)	No / not disclosed
FedRAMP authorized	No / not disclosed
GDPR compliant	Yes
CCPA compliant	Yes
HIPAA compliant	No / not disclosed
NIST AI RMF aligned	No / not disclosed
CSA STAR certified	No / not disclosed
EU AI Act classification	general_purpose

Data handling

Trains on user data	no
Outputs feed model improvement	no
Data retention period	User-controlled
Can delete user data on request	Yes (SLA 30 days)
Default data residency	US
Encryption at rest	Yes (AES-256)
Encryption in transit	Yes
DPA available	Yes
Public subprocessor list	No / not disclosed
HIPAA BAA available	No / not disclosed

IP profile

User owns outputs	yes
Vendor claims output rights	No / not disclosed
Input IP protection	moderate
Indemnification offered	No / not disclosed
Copyright shield program	No / not disclosed
Commercial use permitted	Yes
Training data provenance	partially_disclosed
Known IP lawsuits	No / not disclosed

Jurisdiction

Incorporation country	United States
Incorporation jurisdiction risk	low
Subject to US jurisdiction	Yes
Subject to EU jurisdiction	Yes
Subject to China jurisdiction	No / not disclosed
Subject to Russia jurisdiction	No / not disclosed
Government data access risk	moderate
Five Eyes aligned	Yes
Adequate privacy jurisdiction	No / not disclosed

Governance

Publishes model cards	Yes
Publishes transparency reports	Yes
Has AI ethics board	Yes
Safety testing disclosed	Yes
Red-teaming program	Yes
Government contracts	No / not disclosed
Responsible-AI policy	https://huggingface.co/blog/ethical-charter
Terms of service	https://huggingface.co/terms-of-service
Privacy policy	https://huggingface.co/privacy

Incidents on record

Date	Severity	Incident
2023-06-01	medium	Malicious model upload and pickle deserialization risks [source]

OWASP LLM Top 10 cross-walk

TrustAtlas dimensions that materially address each OWASP risk. Use to translate this vendor's compliance posture and data-handling stance into the application-security vocabulary your security team already uses.

LLM01	Prompt Injection User-supplied prompts manipulate model behaviour to bypass intended controls. SecurityTransparencyDependency chain
LLM02	Sensitive Information Disclosure Models leak PII, PHI, secrets, or proprietary data through outputs. Data handlingIP exposureJurisdiction
LLM03	Supply Chain Risk propagates from upstream models, datasets, plug-ins, and vendors. Dependency chainBusiness stabilitySecurity
LLM04	Data and Model Poisoning Adversarial training data or fine-tuning input degrades model integrity. Data handlingTransparencySecurity
LLM05	Improper Output Handling Downstream systems blindly trust model output, enabling injection downstream. IP exposureTransparency
LLM06	Excessive Agency Agents granted overbroad tool, identity, or permission scopes cause harm. Dependency chainTransparencyJurisdiction
LLM07	System Prompt Leakage System prompts containing secrets or logic are extracted via crafted input. Data handlingTransparency
LLM08	Vector and Embedding Weaknesses Vector stores and RAG pipelines leak or contaminate retrieved context. Data handlingSecurity
LLM09	Misinformation Hallucinated, biased, or fabricated outputs treated as authoritative. TransparencyRegulatory complianceBusiness stability
LLM10	Unbounded Consumption Cost, denial-of-service, and resource-exhaustion attacks against LLM endpoints. SecurityBusiness stability

Full framework reference: https://trustatlas.pages.dev/framework/owasp-llm-top-10

NIST AI RMF cross-walk

How each NIST AI RMF function is supported by the dimensions TrustAtlas scores.

GOVERN	Govern Establish AI governance structure: policies, roles, accountability. Regulatory complianceJurisdictionTransparencyBusiness stability
MAP	Map Establish AI context: intended purpose, use cases, capabilities, and risks. TransparencyDependency chainData handlingIP exposure
MEASURE	Measure Quantitative + qualitative risk assessment: testing, benchmarks, monitoring. SecurityData handlingTransparency
MANAGE	Manage Treat identified risks: mitigation, controls, incident response, lifecycle. Regulatory complianceSecurityDependency chainBusiness stability

Full framework reference: https://trustatlas.pages.dev/framework/nist-ai-rmf

Cited sources

Field	Source
data_handling.data_retention_period	https://huggingface.co/privacy Verified 2026-04-19 by admin
data_handling.trains_on_user_data	https://huggingface.co/terms-of-service Verified 2026-04-19 by admin
ip_profiles.training_data_provenance	https://huggingface.co/docs/hub/model-cards Verified 2026-04-19 by admin
ip_profiles.user_owns_outputs	https://huggingface.co/terms-of-service Verified 2026-04-19 by admin
jurisdiction_profiles.incorporation_country	https://huggingface.co/company Verified 2026-04-19 by admin
security_compliance.gdpr_compliant	https://huggingface.co/privacy Verified 2026-04-19 by admin
security_compliance.soc2_type2	https://huggingface.co/security Verified 2026-04-19 by admin

Questions to ask before signing

Vendor-agnostic baseline. Send these to the vendor and require written answers before contract.

01. Provide your most recent SOC 2 Type II report (with bridge letter if applicable).
02. Describe your training-data provenance and customer opt-out mechanics in writing.
03. List all sub-processors and confirm notification policy for material additions.
04. Confirm BAA availability and signed-BAA process if we process PHI.
05. Describe rate-limiting, quota, and circuit-breaker controls protecting our usage.
06. Provide your model card or equivalent disclosure documenting intended use, limitations, and known failure modes.
07. Describe your prompt-injection defences and red-team posture against OWASP LLM Top 10 risks.
08. Confirm data residency options and which sub-regions our data may touch.
09. Provide incident-response SLAs, security-event notification timelines, and the most recent pen-test report summary.
10. Confirm output ownership terms and any indemnification or copyright-shield programs available.
11. Describe acquisition-risk safeguards and what happens to our data on a change of control.
12. List foundation-model dependencies and how upstream-model risk is mitigated.

Methodology + caveats

Composite scores use the default-balanced weight profile (25% data handling, 20% IP exposure, 15% jurisdiction, 15% security, 10% regulatory compliance, 8% transparency, 5% business stability, 2% dependency chain). All facts are sourced from the vendor's own public disclosures, public regulatory filings, or reputable secondary reporting — see the cited sources table above. This pack is decision-support material, not legal advice or audit evidence.