Open models and an established disclosure process, but repeated serious chip flaws
Overall rating “solid”. No deduction for controversies. The superscript numbers link directly to the source.
1 · Responsibility
AMD has general oversight structures, but the Responsible AI Council and the AI review procedures are documented only as principles without a mandate or results.
What is documented
- The Audit and Finance Committee oversees IT and cybersecurity risks and receives an annual report from an independent adviser1
- The Innovation and Technology Committee is responsible for risk mitigation for products based on new technology1
- The Chief Risk and Compliance Officer reports quarterly; there is annual mandatory training and the anonymous AMD Aware hotline2 5
- The code of conduct requires an AI Use Case Assessment, approved tools and an AI Security Standard for internal AI uses5
- According to an excerpt of the 10-K, the cyber programme is aligned with the NIST AI Risk Management Framework6
The case against
AI governance only as a principle
- In the 2026 proxy statement no committee has an explicit AI mandate, and no person responsible for AI is named1
- The Responsible AI Council is only named; its mandate, chair and reporting line are not disclosed4
- Data and model review, red teaming and vendor assessment are named in the human rights policy only as principles, without results3
Sources for Responsibility (11)
- AMD / U.S. SEC: Proxy Statement DEF 14A 2026 · 13 May 2026 · primary source
- AMD: Ethics and Compliance – Our Approach · 3 November 2026 · primary source
- AMD: Human Rights Policy (Updated August 2026) · 3 November 2026 · secondary source
- AMD: Responsible AI – Our Commitment · 3 November 2026 · secondary source
- AMD: AMD Global Code of Conduct (October 2025) · 3 November 2026 · secondary source
- Board Cybersecurity (extract from 10-K): ADVANCED MICRO DEVICES INC 10-K Item 1C. Cybersecurity – 2026-02-03 · 3 February 2026 · secondary source
- AMD: AMD Responsible AI Use Policy · 3 November 2026 · secondary source
- AMD: 2025–26 Corporate Responsibility Report · 18 August 2026 · primary source
- AMD / U.S. SEC: Form 8-K, voting results of the 2025 annual general meeting · 14 May 2025 · primary source
- AMD Silo AI: AMD Silo AI Signs the EU AI Code of Practice · 20 August 2025 · secondary source
- Freevacy: EU Commission publishes GPAI Code of Practice assessment and signatory list · 1 August 2025 · secondary source
2 · Safety
AMD fixes reported flaws through a documented disclosure process and tests its own models, but repeated serious flaws in chips and firmware stand against this.
What is documented
- AMD runs a PSIRT with coordinated disclosure, CNA status, FIRST membership and a bug bounty via Intigriti4 5
- AMD publishes security bulletins, including for AI software, for example on four flaws in Ryzen AI Software and the NPU driver6
- A fix for EntrySign was shipped to customers; mitigations are available for Sinkclose9 8
- In-house models are evaluated with ToxiGen, CrowS-Pairs, TruthfulQA, adversarial tests and red teaming1
- Against diversion there is an anti-diversion committee, customer screening, locking software and cooperation with Conflict Armament Research14
The case against
Repeated serious flaws
- Via LeftoverLocals, LLM responses could be reconstructed from GPU memory; at disclosure AMD was still assessing mitigations, and their status is unclear7
- Sinkclose, EntrySign and RMPocalypse affected processor firmware and confidential computing environments; no exploitation is documented8 9 10
- According to press reports, a flaw in AMD AutoUpdate remained open for up to 124 days; AMD refused the bounty payment11
- The bug bounty programme does not cover ROCm and Instinct5
Sources for Safety (16)
- AMD Developer (E. Barsoum et al.): Building Trust Through Open Innovation · 23 October 2025 · primary source
- AMD / Hugging Face: Model card Instella-3B-Instruct · 3 November 2026 · primary source
- Board Cybersecurity (extract from 10-K): ADVANCED MICRO DEVICES INC 10-K Item 1C. Cybersecurity – 2026-02-03 · 3 February 2026 · secondary source
- AMD: Product Security (PSIRT) · 3 November 2026 · primary source
- Intigriti: AMD Bug Bounty Program – Detail · 3 November 2026 · primary source
- AMD PSIRT: AMD-SB-7037: AMD Ryzen AI Software Vulnerabilities · 2 April 2025 · primary source
- Trail of Bits: LeftoverLocals: Listening to LLM responses through leaked GPU local memory · 16 January 2024 · primary source
- CERT-EU: Security Advisory 2024-075: Vulnerabilities in AMD CPUs · 12 August 2024 · primary source
- oss-security (Openwall), Google disclosure: AMD Zen microcode signature verification vulnerability (CVE-2024-56161) · 6 March 2025 · primary source
- SecurityWeek: RMPocalypse: New Attack Breaks AMD Confidential Computing · 3 November 2026 · secondary source
- BGR: AMD Refuses To Pay A $10,000 Bug Bounty For A Flaw It Never Caught · 15 June 2026 · secondary source
- Mixed News: AMD's Instinct security bulletin names two CVEs for one flaw, and explains only one · 6 October 2026 · secondary source
- AMD: AMD Responsible AI Use Policy · 3 November 2026 · secondary source
- AMD: 2025–26 Corporate Responsibility Report · 18 August 2026 · primary source
- SecurityWeek: AMD Investigating Breach Claims After Hacker Offers to Sell Data · 19 June 2024 · secondary source
- AMD: AMD Global Code of Conduct (October 2025) · 3 November 2026 · secondary source
3 · Transparency
AMD discloses the weights, training data and code of its own models and acknowledges limitations, but documents no labelling of AI-generated content and no AI transparency report.
What is documented
- For Instella, AMD OLMo, Hummingbird and Zebra-Llama, weights, training data, configurations and code are open6
- The Instella-3B-Instruct model card lists datasets for each training stage with token counts and licences5
- The model card contains a section on bias, risks and limitations and excludes areas of use5
- AMD acknowledges limitations regarding jailbreaks and factual accuracy6
- The 10-K names risks arising from the responsible use of AI4
The case against
Labelling and reporting
- No information was found on labelling AI-generated content from the in-house image and video models
- An AI transparency report is missing
- The Instella model card contains no safety or bias benchmarks5
Unclear disclosure
- Bulletin AMD-SB-6033 on the AI library RCCL lists two CVE numbers but explains only one; the fix came about eleven weeks earlier3
- The model cards for Poro and Viking from Silo AI were not reviewed
Sources for Transparency (6)
- AMD: AMD Responsible AI Use Policy · 3 November 2026 · secondary source
- AMD PSIRT: AMD-SB-7037: AMD Ryzen AI Software Vulnerabilities · 2 April 2025 · primary source
- Mixed News: AMD's Instinct security bulletin names two CVEs for one flaw, and explains only one · 6 October 2026 · secondary source
- AMD / U.S. SEC: Form 10-K for the fiscal year ended 2025-12-27 · 4 February 2026 · primary source
- AMD / Hugging Face: Model card Instella-3B-Instruct · 3 November 2026 · primary source
- AMD Developer (E. Barsoum et al.): Building Trust Through Open Innovation · 23 October 2025 · primary source
4 · Why the harm rule does not apply
The harm rule does not apply: within the window there is no judgment, regulatory decision or fine concerning harm caused by AI. The recorded flaws are vulnerabilities without documented harm.
5 · Controversies (5)
- mediumRecurring serious flaws in processor firmware and confidential computing: Sinkclose (CVE-2023-31315, CVSS 7.5), EntrySign for forging microcode (CVE-2024-56161, public on 2025-02-03) and RMPocalypse against SEV-SNP (CVE-2025-0033, 2025); in each case without documented exploitation. · 9 August 2024 · unklar
- mediumTrail of Bits disclosed the LeftoverLocals flaw (CVE-2023-4969), through which other users of an AMD GPU can reconstruct LLM responses of other processes from local GPU memory; AMD was still affected at disclosure and was still assessing mitigations. · 16 January 2024 · unklar
- minorBulletin AMD-SB-6033 on a flaw in the AI communication library RCCL for Instinct accelerators lists two CVE numbers but explains only one; the fix came about eleven weeks before publication. · 30 September 2026 · unklar
- minorA flaw in AMD AutoUpdate (CVE-2026-40677) allowed code execution via man-in-the-middle and, according to press reports, remained open for up to 124 days; AMD refused the bounty payment and asked for the blog post to be removed. · 12 June 2026 · unklar
- minorA hacker allegedly offered data stolen from AMD or a hosting partner on future products, employees, customers and source code; AMD confirmed an investigation with authorities, but no outcome is documented. · 18 June 2024 · unklar
6 · What remains open
- Board responsibility for AI and composition of the Responsible AI Council
- Evidence of how the AI Use Case Assessment is implemented
- Remediation status of the LeftoverLocals flaw
- Item 1C and Item 3 of the 10-K and the governance chapter of the CR report could not be read in the original
- Labelling of AI-generated content from in-house generative models
- Enforcement of the AI usage policy and reporting channel for misuse
7 · How it was assessed
Cut-off date: 3 November 2026 · 45 searches, 39 fetches.
Primary documents read (10)
- Form 10-K FY2025 (S03, two retrievals, Part 1 through Risk Factors; Item 1C/3 not included)
- Form 10-K/A FY2025 (S04)
- Form 8-K Voting Results 2025 (S05)
- Proxy Statement DEF 14A 2026 (S06, two retrievals, complete)
- 2025–26 Corporate Responsibility Report (S11, two retrievals through p. 52)
- Responsible AI page (S08) and Responsible AI Use Policy (S09)
- Global Code of Conduct Oct. 2025 (p.12), Ethics-and-Compliance page (p.13), Human Rights Policy Aug. 2026 (p.14)
- PSIRT page (S15), Bug Bounty Program (S16), Bulletin AMD-SB-7037 (S20)
- Model card Instella-3B-Instruct (S19) and developer contribution to security benchmarks (S18)
- Official and independent findings: CERT-EU 2024-075 (S22), Google EntrySign disclosure (S23), Trail of Bits LeftoverLocals (S24)
Research of 3 November 2026 under Globalance AI Footprint rulebook version 0