← Back to all companies
D

DeepSeek

Chinese AI lab producing competitive open-source models like DeepSeek-R1 and V3.

Safety Documents

❌
Responsible Scaling Policy
No publicly available Responsible Scaling Policy or equivalent safety framework document found. Stanford FMTI transparency report for DeepSeek (December 2025) available.🔗
⚠️
Model Card / System Card
DeepSeek-R1 Technical Report / Model Documentation. Technical reports available on Hugging Face and GitHub. Not in the same format as standard model cards with explicit safety evaluation sections.🔗
⚠️
Safety Benchmark Results
Technical reports include some benchmark results but detailed safety evaluation results are not publicly published in the format of other frontier labs.🔗
✅
Acceptable Use Policy
DeepSeek has Terms of Service and Model License Agreement. Specific AUP less comprehensively documented than US/EU counterparts.🔗

Testing & Evaluation

✅
Third-Party Red Teaming
KELA Cyber (independent); Qualys TotalAI (independent); Adversa AI (independent). Multiple third-party security firms independently red-teamed DeepSeek R1 and found significant vulnerabilities. These were NOT commissioned by DeepSeek. DeepSeek itself has not published a third-party commissioned red-team report.🔗
⚠️
CBRN Risk Evaluation
No published CBRN-specific evaluation found. External researchers found DeepSeek R1 could be prompted to provide bioweapon instructions. TechCrunch reported R1 produced 'plans for a bioweapon attack' under adversarial prompting.
⚠️
Pre-Deployment Safety Evaluation
No published pre-deployment safety evaluation found in the format used by other frontier labs.
❌
External Safety Evaluations
No evidence of participation in METR, ARC, UK AISI, or other formal external evaluation programs.

Governance

❌
Independent Safety Board
No publicly documented independent safety board or advisory committee found.
❌
Seoul AI Safety Commitment
DeepSeek was not among the 16 companies that signed the Seoul Frontier AI Safety Commitments in May 2024. Company was not founded at sufficient scale at that time.🔗
❌
Government Safety Report
No evidence of voluntary safety report submission to any government. Multiple governments (Italy, Australia, Taiwan) have restricted or investigated DeepSeek over privacy concerns.🔗
❌
Third-Party Audits
No evidence of commissioned third-party safety audits.
⚠️
Whistleblower Policy
No publicly documented whistleblower or safety concern process found.

Policy Positions

Military Use:unknown
Model license agreement restricts some uses but explicit military use policy not prominently documented. Chinese company — subject to Chinese national security laws.🔗
Surveillance Use:unknown
No explicit surveillance policy found. Multiple Western governments have raised concerns about potential Chinese government data access.
Open-Source Models:Yes
DeepSeek-V2, DeepSeek-V3, DeepSeek-V3.2, DeepSeek-R1, DeepSeek-R1-Zero. and 4 more. Most major models released under MIT License or DeepSeek Model License. R1 and V3 under MIT License.🔗
Children/Minors Policy:Unknown
No publicly documented children/minors-specific policy found.

Incident History

Wiz Research Uncovers Exposed Database with 1M+ Lines of Sensitive Data

2025-01-29

Security firm Wiz discovered a publicly accessible ClickHouse database linked to DeepSeek exposing over 1 million lines of sensitive data including user chat histories, API keys, and backend operational details. Ports 8123 and 9000 were open to the internet.

🔗 Source🔗 Company Response

DeepSeek R1 Fails Jailbreak Tests — 100% Attack Success Rate Reported

2025-01-27

Multiple security researchers (KELA, Qualys, Adversa AI, HarmBench researchers) found DeepSeek R1 was highly vulnerable to jailbreak attacks. One study reported a 100% attack success rate on 50 HarmBench prompts. Model produced bioweapon instructions, explosive device guides, and self-harm promotion content.

🔗 Source🔗 Company Response

Italy and Multiple Governments Restrict or Investigate DeepSeek

2025-01-30

Italy's data protection authority restricted DeepSeek AI from processing Italian users' data, citing concerns over privacy and data storage in China. Australia, Taiwan, and other governments took similar steps.

🔗 Source🔗 Company Response

CNN/CCDH Investigation: DeepSeek Ended Violence-Planning Conversation with 'Happy (and safe) shooting!'

2026-03-11

A CNN investigation with the Center for Countering Digital Hate (CCDH) found that when a teenage test user asked DeepSeek for information that could be used in an attack on Irish opposition leader Mary Lou McDonald, the chatbot ended the conversation by wishing the user 'Happy (and safe) shooting!' The report found DeepSeek among the most egregious examples of AI chatbots failing to prevent violence-planning assistance.

🔗 Source🔗 Company Response

Timeline

2026-03-11
CNN/CCDH investigation highlights DeepSeek among worst performers on violence-planning content moderation, including chatbot wishing user 'Happy (and safe) shooting!'🔗
2025-01-30
Italy restricts DeepSeek; multiple governments launch investigations🔗
2025-01-27
Multiple security firms report R1 jailbreak vulnerabilities; Wiz discovers exposed database🔗
2025-01-20
DeepSeek-R1 released under MIT License, achieving reasoning performance competitive with OpenAI o1🔗
2024-12-26
DeepSeek-V3 released under MIT License, competitive with GPT-4 class models🔗