Tested in-house models and cloud services, pending lawsuits over facial recognition
Overall rating “solid”. No deduction for controversies. The superscript numbers link directly to the source.
1 · Responsibility
Amazon has board oversight of AI and a safety framework for frontier models, but discloses no due diligence for government and military customers.
What is documented
- The Nominating and Corporate Governance Committee oversees risks from responsible AI development and AI governance2 3
- A Security Committee of independent directors is responsible for security; the Chief Security Officer reports to the CEO1
- The Frontier Model Safety Framework defines thresholds for CBRN, cyber and automated AI research and provides for a go/no-go decision by senior leadership4
- Bedrock, Q Business, Textract and Transcribe are certified to ISO/IEC 42001; the 2025 surveillance audit found nothing10
- A central Responsible AI Governance Team is described in the report under the G7 Hiroshima framework5
The case against
Limited scope
Government customers and shareholders
- Investigations accuse Amazon of accepting clauses in the Project Nimbus contract with Israel that undermine its own terms of use8
- No review of the use of AI and cloud services by military and security agencies is disclosed9
- Two AI proposals were rejected in 2025; a proposal on the Responsible AI Approach was not put to a vote in 202613 15
Sources for Responsibility (15)
- Amazon.com, Inc. (SEC-Einreichung): Form 10-K for fiscal year 2025 · 3 November 2026 · primary source
- Amazon.com, Inc.: 2026 Proxy Statement (PDF) · 3 November 2026 · primary source
- Amazon.com, Inc. (SEC Filing): DEF 14A 2026 · 3 November 2026 · primary source
- Amazon.com, Inc. (Amazon Science): Amazon's Frontier Model Safety Framework · 3 November 2026 · primary source
- OECD.AI (Report by Amazon): G7 Hiroshima AI Process Reporting Framework – Amazon (Nova models on Bedrock) · 31 October 2025 · primary source
- Amazon Web Services: AWS Responsible AI Policy · 13 January 2025 · secondary source
- Ranking Digital Rights: 2025 Big Tech Edition – Amazon · 13 May 2025 · primary source
- Democracy Now!: Report: Google and Amazon violated their own terms of service in Israel's $1.2 billion Project Nimbus deal · 31 October 2025 · secondary source
- Intercommunity Peace & Justice Center (Aktionärsantrag): Shareholder proposal on aligning AI sales and use with the Responsible AI Approach · 3 November 2026 · secondary source
- Amazon Web Services (AWS Security Blog): AWS successfully completed its first surveillance audit for ISO 42001:2023 with no findings · 26 February 2026 · primary source
- Amazon.com, Inc. (About Amazon): Responsible AI – Our approach · 3 November 2026 · secondary source
- Business & Human Rights Resource Centre: US corporate human rights index – Amazon (findings table) · 1 August 2025 · secondary source
- Amazon.com, Inc. (SEC-Einreichung): Form 8-K, voting results of the 2025 annual general meeting · 21 May 2025 · primary source
- Interfaith Center on Corporate Responsibility (ICCR): Amazon 2025 Shareholder Proposals · 23 May 2025 · secondary source
- Investor Advocates for Social Justice (SEC PX14A6G): Notice of Exempt Solicitation for the 2026 annual general meeting · 3 November 2026 · secondary source
2 · Safety
Amazon documents extensive internal and external testing of its Nova models, but a compromised AI assistant, a confirmed privacy breach and biometrics lawsuits stand against this.
What is documented
- The Nova technical report cites more than 300 internal attack techniques and external red-teaming by ActiveFence, Deloitte, CMU and Nemesys3
- The service card documents frontier testing with METR; PRISM Eval and ActiveFence tested Nova Premier independently2 1
- Model weights are protected with encryption and isolation; vulnerabilities can be reported via a bug bounty6
- Bedrock detects abuse automatically, reports CSAM to NCMEC and can suspend accounts11
- For Familiar Faces, Ring cites opt-in, deletion periods and internal privacy impact assessments14
The case against
Incidents with in-house AI tools
- A deletion command reached version 1.84.0 of the Amazon Q extension via an overly broad GitHub token; AWS has fixed the flaw8 9
- Media report an AWS outage after changes made by the agent tool Kiro; Amazon cites user error4
- Amazon Quick lacked an authorisation check, so blocked AI agents were reachable; the gap was fixed10
Privacy and biometrics
- The Conseil d'État upheld a breach of data minimisation in performance data of warehouse workers and reduced the fine to EUR 15 million18
- According to the lawsuit, Ring Familiar Faces captures facial data of visitors and passers-by without consent; the case is pending16
- A class action under the Illinois biometrics law has been certified over virtual try-on; liability has not been decided17
Sources for Safety (21)
- Amazon.com, Inc. (Amazon Science): Independent evaluations demonstrate Nova Premier's safety · 29 May 2025 · primary source
- Amazon Web Services: AWS AI Service Card: Amazon Nova Micro, Lite, Pro, and Premier · 3 November 2026 · primary source
- Amazon AGI (arXiv): The Amazon Nova Family of Models: Technical Report and Model Card · 17 March 2025 · primary source
- The Register: Amazon's vibe-coding tool Kiro reportedly vibed too hard and brought down AWS · 20 February 2026 · secondary source
- Fortune: Amazon retail site outages and AI agent inaccurate advice · 12 March 2026 · secondary source
- Amazon.com, Inc. (Amazon Science): Amazon's Frontier Model Safety Framework · 3 November 2026 · primary source
- Amazon.com, Inc. (SEC-Einreichung): Form 10-K for fiscal year 2025 · 3 November 2026 · primary source
- Amazon Web Services: Security Bulletin AWS-2025-015: Amazon Q Developer Extension for Visual Studio Code (Version #1.84) · 23 July 2025 · primary source
- BleepingComputer: Amazon AI coding agent hacked to inject data wiping commands · 24 July 2025 · secondary source
- Pulse (via Fog Security): Amazon Quick authorisation bypass let users reach blocked AI chat agents · 12 May 2026 · secondary source
- Amazon Web Services: Amazon Bedrock abuse detection (user guide) · 3 November 2026 · primary source
- Amazon Web Services: AWS Responsible AI Policy · 13 January 2025 · secondary source
- ConductAtlas (via AWS Service Terms): Amazon Rekognition police department moratorium · 3 November 2026 · secondary source
- Ring LLC (Brief an Senator Markey, veröffentlicht durch das Büro Markey): Ring's reply to the letter of 30 October 2025 on Familiar Faces · 21 November 2025 · primary source
- Electronic Frontier Foundation: The Legal Case Against Ring's Face Recognition Feature · 3 November 2025 · secondary source
- U.S. District Court, Western District of Washington (CourtListener): Sigwalt v. Amazon.com, Inc. and Ring LLC, No. 2:26-cv-01887, complaint · 1 June 2026 · secondary source
- U.S. Court of Appeals for the Seventh Circuit: Svoboda v. Amazon.com Inc., No. 25-1361 (class certification ruling) · 17 December 2025 · primary source
- Conseil d'État (France): Décision n° 492830, Amazon France Logistique · 23 December 2025 · primary source
- Actu-Juridique: La CNIL sanctionne Amazon pour utilisation d'un système de surveillance excessivement intrusif · 23 January 2024 · secondary source
- The Register: Amazon to kill off local Alexa processing, all voice requests shipped to the cloud · 17 March 2025 · secondary source
- Tom's Hardware: Amazon's Rufus AI shopping assistant can be easily jailbroken · 9 March 2026 · secondary source
3 · Transparency
Amazon publishes model cards and service cards and Seals generated images and videos, but describes training data only by category.
What is documented
- For the Nova models there is a technical report with a model card and AWS AI Service Cards stating limits of use1 5
- Amazon has published a report under the G7 Hiroshima framework on the Nova models2
- Outputs of Nova Canvas, Reel and Sonic carry invisible watermarks; there is a detection tool for Titan images1 3
- The service cards recommend human review for consequential decisions and disclosure to end users5
- Amazon has signed the EU code of practice for general-purpose AI models4
The case against
Training data
Sources for Transparency (8)
- Amazon AGI (arXiv): The Amazon Nova Family of Models: Technical Report and Model Card · 17 March 2025 · primary source
- OECD.AI (Report by Amazon): G7 Hiroshima AI Process Reporting Framework – Amazon (Nova models on Bedrock) · 31 October 2025 · primary source
- Amazon Web Services: Watermark detection for Amazon Titan Image Generator now available in Amazon Bedrock · 23 April 2024 · primary source
- Euronews: Some 26 tech companies sign up to EU Commission's AI code · 1 August 2025 · secondary source
- Amazon Web Services: AWS AI Service Card: Amazon Nova Micro, Lite, Pro, and Premier · 3 November 2026 · primary source
- Ranking Digital Rights: 2025 Big Tech Edition – Amazon · 13 May 2025 · primary source
- GeekWire: Amazon sued by YouTubers for allegedly scraping their content to train AI video tool · 3 November 2026 · secondary source
- Ring LLC (Brief an Senator Markey, veröffentlicht durch das Büro Markey): Ring's reply to the letter of 30 October 2025 on Familiar Faces · 21 November 2025 · primary source
4 · Why the harm rule does not apply
The harm rule does not apply: the AI link of the fine upheld in France is not established, and the other cases are pending proceedings without findings.
5 · Controversies (10)
- mediumA class action in the federal court in Seattle accuses Amazon and Ring of capturing facial data of visitors and passers-by without consent through the AI feature Familiar Faces; Senator Markey and the EFF had previously criticised the feature. · 1 June 2026 · open
- mediumIn December 2023 the CNIL fined Amazon France Logistique EUR 32 million for excessive monitoring of warehouse workers via scanner indicators and video surveillance; the Conseil d'État overturned the objection to the indicators but upheld the breach of data minimisation (31-day retention) and reduced the fine to EUR 15 million. · 23 December 2025 · abgeschlossen
- mediumThe US Court of Appeals for the Seventh Circuit upheld certification of a class action under the Illinois biometrics law against Amazon's virtual try-on, which captures facial geometry; liability has not been decided. · 17 December 2025 · open
- mediumAn investigation by +972 Magazine, Local Call and The Guardian accuses Amazon and Google of accepting clauses in the Project Nimbus contract with Israel that undermine their own terms of use; the cloud and AI services are said to be used in the Gaza war. · 31 October 2025 · unklar
- mediumVia an overly broad GitHub token, an attacker inserted a deletion command into the Amazon Q Developer extension for VS Code, which was shipped as version 1.84.0 (CVE-2025-8217); according to AWS, execution failed because of a syntax error. · 17 July 2025 · abgeschlossen
- mediumYouTube channel operators are suing Amazon for allegedly using millions of videos to train Nova Reel by circumventing YouTube's protective measures (DMCA); Amazon did not comment. · · open
Further controversies (4)
- minorInvestors criticised Amazon for not putting a 2026 proposal on aligning AI sales (including to security agencies) with its own Responsible AI Approach to a vote and called for votes against four directors; in 2025 two AI proposals were rejected with around 10% support. · 20 May 2026 · abgewiesen
- minorAmazon Quick lacked server-side authorisation, so users could reach blocked AI chat agents; according to trade press, Rufus could easily be led into off-topic answers. · 12 March 2026 · abgeschlossen
- minorAmazon discontinued the option for Echo devices to process voice recordings locally in order to enable generative AI features of Alexa+; since then all requests have been processed in the cloud. · 28 March 2025 · unklar
- minorAccording to the Financial Times (as reported by The Register), the agent tool Kiro caused a 13-hour outage of an AWS service after deleting and recreating an environment; in March 2026 incorrect advice from an AI agent contributed to an outage of the online store. · · unclear
6 · What remains open
- Full list of AI-related legal proceedings; Note 7 of the 10-K and the Q3 2026 10-Q could not be read
- Remedial measures following the decision of the CNIL and the Conseil d'État
- Disclosed human rights review of AI sales to government agencies and the military
- Group-wide figures on AI incidents and complaints
- No stand-alone responsible AI report from Amazon was found
7 · How it was assessed
Cut-off date: 3 November 2026 · 42 searches, 47 fetches.
Primary documents read (17)
- Form 10-K FY 2025 (S1, characters 0–117,795)
- Form 10-Q Q2 2026 (S2, characters 0–100,000)
- 2026 Proxy Statement PDF (S3, characters 0–132,531)
- DEF 14A 2026 (S4, characters 0–100,000)
- Form 8-K Voting Results 2025 (S5)
- Frontier Model Safety Framework (S9)
- AWS AI Service Card Nova (S11)
- Nova Technical Report and Model Card (S12)
- HAIP Report Amazon (S14)
- AWS Responsible AI Policy (S15)
- Bedrock abuse detection (S16)
- Security Bulletin AWS-2025-015 (S26)
- Ring Letter to Senator Markey (S31)
- Complaint Sigwalt v. Amazon (S33)
- Judgment Svoboda v. Amazon, 7th Cir. (S35)
- Conseil d'État 492830 (S24)
- 2025 Amazon Sustainability Report (S20, characters 0–125,567)
Research of 3 November 2026 under Globalance AI Footprint rulebook version 0