Well-documented chip and system security, but AI rules only announced
Overall rating “solid”. No deduction for controversies. The superscript numbers link directly to the source.
1 · Responsibility
MediaTek has documented general governance with an Audit Committee and an information security committee, but its AI framework, usage rules and audit mechanism are still only planned measures.
What is documented
- The Audit Committee oversees risks; internal audit reports directly to the board2 4
- Five of the nine board members are independent, one of them with recognised AI expertise2 4
- An Information Security Committee chaired by the Co-COO and the CFO reports annually to the board and discusses the security of AI applications1
- A confidential reporting channel leads to the Chief Audit Executive; for 2024 MediaTek reports no material legal violations and no fines in two years4 1
- An AI Training Committee exists in the talent function5
The case against
Intentions only
Missing accountability
- A published AI policy, a risk classification and a named responsibility for AI risks are missing
- There is no information on AI training, no external audit of AI processes and no AI-related complaints channel
Sources for Responsibility (7)
- MediaTek Inc.: 2024 MediaTek Sustainability Report, English version, chapter 4 (Corporate Governance, chapter PDF) · 3 November 2026 · primary source
- MediaTek Inc.: 2025 English Annual Report · 3 November 2026 · primary source
- 聯發科技股份有限公司: MediaTek Annual Report 2025 (Chinese version) · 3 November 2026 · primary source
- MediaTek Inc.: Corporate Governance (Investor Relations) · 3 November 2026 · primary source
- MediaTek Inc.: MediaTek 2025 ESG Report (English) · 3 November 2026 · primary source
- MediaTek Inc.: 2024 MediaTek ESG Report (English, full document) · 3 November 2026 · primary source
- MediaTek DaVinci (documentation wiki dvcbot.atlassian.net): MediaTek DaVinci FAQ (GAI Service Framework) · 3 November 2026 · secondary source
2 · Safety
MediaTek documents its security organisation with certificates, red-team exercises and a public vulnerability policy, but there is no documented abuse monitoring for the publicly usable models.
What is documented
- ISO 27001 certification, TISAX and product certificates under ISO/SAE 21434; a zero-trust architecture since 20243 1
- A CISO, an Information Security Committee, red-team exercises and around 306 information security specialists in 20251 3
- A public vulnerability policy with a 90-day target and monthly bulletins giving device manufacturers at least two months' notice9 10
- The outside view from UpGuard gives grade A, but only for web and e-mail configuration8
- Breeze Guard detects fraud and phishing content in the Taiwanese context and was tested with its own benchmark5 12
- The internal AI factory runs on premises; logging is switched off for Azure OpenAI6 7
The case against
Gaps in AI abuse protection
- Neither usage restrictions nor abuse monitoring are documented for the publicly usable Breeze models and the LINE service
- Breeze-7B was released without particular handling of harmful content4
- Testing and monitoring of the AI agents used internally are not documented
Incidents and gaps in the evidence
- The chip vulnerability CVE-2025-20435 allowed encrypted data to be read on certain smartphones; the roll-out of the fix depends on the manufacturers13
- The scope of the ISO 27001 certificate is not stated; the chapter on information security in the 2025 report remained unread
Sources for Safety (13)
- MediaTek Inc.: MediaTek 2025 ESG Report (English) · 3 November 2026 · primary source
- MediaTek Inc.: 2024 MediaTek ESG Report (English, full document) · 3 November 2026 · primary source
- MediaTek Inc.: 2024 MediaTek Sustainability Report, English version, chapter 4 (Corporate Governance, chapter PDF) · 3 November 2026 · primary source
- MediaTek Research (Hugging Face): Model Card Breeze-7B-Instruct-v1_0 · 5 March 2024 · primary source
- MediaTek Research / arXiv: Taiwan Safety Benchmark and Breeze Guard: Toward Trustworthy AI for Taiwanese Mandarin (arXiv 2603.07286v1) · 3 November 2026 · primary source
- NVIDIA: MediaTek AI Factory (Customer Story) · 11 December 2025 · secondary source
- MediaTek DaVinci (documentation wiki dvcbot.atlassian.net): MediaTek DaVinci FAQ (GAI Service Framework) · 3 November 2026 · secondary source
- UpGuard: MediaTek Security Rating · 3 November 2026 · primary source
- MediaTek Inc.: Vulnerability Disclosure Policy · 3 November 2026 · primary source
- MediaTek Inc.: Product Security Bulletin December 2025 · 1 December 2025 · primary source
- MediaTek Inc.: Privacy Policy · 22 July 2026 · secondary source
- MediaTek Inc. (Tek Talk Blog): MediaTek Research Breeze 3: 讓 AI 聽懂台語、說出台味、守護台灣 · 26 February 2026 · secondary source
- IT Brief Australia: Critical MediaTek flaw lets attackers steal phone crypto · 12 March 2026 · secondary source
3 · Transparency
MediaTek publishes model cards and discloses method and limitations in the Breeze Guard paper, but training data and limitations of the other models remain vague.
What is documented
- The model cards for Llama-Breeze2 and Breeze-7B state base model, architecture, licence and benchmarks6 7
- The Breeze Guard paper discloses false alarms, the restriction to single conversation turns, synthetic data and an incomplete taxonomy8
- The sustainability reports state openly that an AI governance framework does not yet exist2 3
- Monthly product security bulletins systematically disclose vulnerabilities with CVE details4
The case against
Thin model information
Sources for Transparency (10)
- MediaTek Inc.: MediaTek 2025 ESG Report (English) · 3 November 2026 · primary source
- MediaTek Inc.: 2024 MediaTek ESG Report (English, full document) · 3 November 2026 · primary source
- MediaTek Inc.: 2024 MediaTek Sustainability Report, English version, chapter 4 (Corporate Governance, chapter PDF) · 3 November 2026 · primary source
- MediaTek Inc.: Product Security Bulletin December 2025 · 1 December 2025 · primary source
- MediaTek Inc.: Privacy Policy · 22 July 2026 · secondary source
- MediaTek Research (Hugging Face): Model Card Llama-Breeze2-8B-Instruct · 23 January 2025 · primary source
- MediaTek Research (Hugging Face): Model Card Breeze-7B-Instruct-v1_0 · 5 March 2024 · primary source
- MediaTek Research / arXiv: Taiwan Safety Benchmark and Breeze Guard: Toward Trustworthy AI for Taiwanese Mandarin (arXiv 2603.07286v1) · 3 November 2026 · primary source
- MediaTek Inc. (Tek Talk Blog): MediaTek Research Breeze 3: 讓 AI 聽懂台語、說出台味、守護台灣 · 26 February 2026 · secondary source
- MediaTek Inc.: Artificial Intelligence (technology page) · 3 November 2026 · secondary source
4 · Why the harm rule does not apply
The harm rule does not apply: in the period examined there is neither a judgment nor an authority decision nor a fine nor a pattern of serious incidents involving its own AI systems.
5 · Controversies (1)
6 · What remains open
- Whether an AI policy has since been adopted and who is responsible for AI risks in the group
- Implementation status of the announced audit mechanism for AI agents and the new usage rules for AI tools
- Scope of the ISO 27001 certification and incident and data protection figures for 2025
- Terms of use, abuse monitoring and reporting channels for the Breeze models and the LINE service
- Chapters on information security, risks and proceedings in the 2025 reports, which were truncated on retrieval
7 · How it was assessed
Cut-off date: 3 November 2026 · 44 searches, 34 fetches.
Primary documents read (11)
- MediaTek 2025 ESG Report, English (to approx. p. 035, two retrievals)
- 2025 Sustainability Report, English, FinancialFilings (abridged web version, two retrievals)
- 2024 MediaTek ESG Report, English (to approx. p. 036)
- 2024 ESG Report, Chapter 4 Corporate Governance (chapter PDF)
- 2025 English Annual Report (pp. 1–34)
- 114年聯發科技年報 (S. 1–27)
- Corporate Governance page, Vulnerability Disclosure Policy, Product Security Bulletin December 2025, Privacy Policy
- Model cards Llama-Breeze2-8B-Instruct and Breeze-7B-Instruct, Breeze-Guard paper (arXiv 2603.07286)
- 股東會議事錄 2026
- MediaTek v Huawei [2025] EWHC 649 (Pat) (to para. 77)
- NBIM exclusion list
Research of 3 November 2026 under Globalance AI Footprint rulebook version 0