Open model cards, but officially established failings in Grok's image tool
Overall rating “limited”. Harm rule applied. The superscript numbers link directly to the source.
1 · Responsibility
A risk framework with thresholds is documented, but the regulator found a late impact assessment, and there is no AI or safety committee.
What is documented
- The 2025 Risk Management Framework contains risk categories, benchmarks and quantitative release thresholds for bio/chem and honesty2
- The 2025 framework provides for anonymous internal reporting and whistleblower protection2
- The acceptable use policy names a reporting channel and the reporting of child sexual abuse material to NCMEC13
- X and xAI committed to the Canadian regulator to provide quarterly reports and annual third-party audits12
The case against
Risk assessment late or incomplete
- According to the Canadian regulator, the impact assessment for Grok Imagine was updated only months after launch; xAI rated sexualised deepfakes as a low risk12
- The 2026 Frontier AI Framework names neither roles nor bodies nor thresholds3 7
- The European Commission, the Irish data protection authority and the ICO are examining whether risks were assessed before Grok features were introduced9 8 11
Sources for Responsibility (13)
- Space Exploration Technologies Corp. (SEC): Prospectus (Form 424B4) · 12 June 2026 · primary source
- xAI: xAI Risk Management Framework · 20 August 2025 · primary source
- xAI: xAI Frontier Artificial Intelligence Framework · 30 June 2026 · secondary source
- Future of Life Institute: AI Safety Index Summer 2026 · 3 November 2026 · primary source
- The Next Web: Sequoia's Roelof Botha joins the SpaceX board, days after the record IPO · 18 June 2026 · secondary source
- Eco-Business: SpaceX's rock bottom ESG rating reignites debate over sustainability scores · 23 June 2026 · primary source
- regulations.ai: xAI Frontier Artificial Intelligence Framework (tracker entry RAIC-XAI-XAIFRON-2026) · 30 June 2026 · secondary source
- The Register: Ireland joins regulator smackdown after X's Grok AI accused of undressing people · 17 February 2026 · primary source
- European Commission: Commission investigates Grok and X's recommender systems under the Digital Services Act · 26 January 2026 · primary source
- PPC Land: FTC orders seven AI chatbot companies to detail child safety measures · 10 September 2025 · primary source
- Information Commissioner's Office (ICO): ICO announces investigation into Grok · 3 February 2026 · primary source
- Office of the Privacy Commissioner of Canada: PIPEDA Findings #2026-004: Investigation into X Corp. and X.AI LLC (Grok) · 11 June 2026 · primary source
- xAI: Acceptable Use Policy · 3 November 2026 · secondary source
2 · Safety
The pillar cannot be better than “limited” because of a serious, officially established incident involving Grok Imagine; on top of this come failures in reliability and system protection.
What is documented
- The Grok 4.20 and 4.6 model cards measure refusals, jailbreaks, CBRN, cyber and honesty, with external testers and a third-party audit on deception6 2
- The Grok 4.6 model card discloses regressions on some safety measurements2
- The acceptable use policy prohibits undressing real people and sexualising children16
- Since January 2026 Grok has blocked editing images of real people to undress them; according to the regulator, effectiveness has not been demonstrated17 18
The case against
Serious incident with image AI
- The Canadian regulator found that X and xAI launched Grok Imagine without adequate safeguards and that the response was insufficient and untimely18
- As early as August 2025, the “Spicy” mode produced a video in a test in which Taylor Swift undresses without an explicit prompt19
- Ofcom, California and the Paris judiciary are investigating; plaintiffs from Tennessee accuse Grok of creating sexual images of them as minors12 13 15 14
Reliability and system protection
- After a code update Grok spread antisemitic answers for about 16 hours; an unauthorised prompt change led to statements about a “white genocide”1 8
- A private API key was public for about two months and allowed access to more than 60 models11
- Common Sense Media considers Kids Mode ineffective and the handling of mental health crises dangerous4
Sources for Safety (19)
- Engadget: Grok team apologizes for the chatbot's 'horrific behavior' and blames 'MechaHitler' on a bad update · 12 July 2025 · primary source
- SpaceXAI (DBA of XAI LLC): Grok 4.6 Model Card · 12 August 2026 · primary source
- Future of Life Institute: AI Safety Index Summer 2026 · 3 November 2026 · primary source
- Common Sense Media: Grok AI Chatbot Not Safe for Teens, Common Sense Media Report Finds · 27 January 2026 · primary source
- Fortune: Elon Musk released xAI's Grok 4 without any safety reports—despite calling AI more 'dangerous than nukes' · 17 July 2025 · primary source
- xAI: Grok 4.20 System Card · 7 April 2026 · primary source
- TechCrunch: Secretaries of state urge X to stop its Grok chatbot from spreading election misinformation · 5 August 2024 · primary source
- TechCrunch: xAI blames Grok's obsession with white genocide on an 'unauthorized modification' · 15 May 2025 · primary source
- xAI: xAI Frontier Artificial Intelligence Framework · 30 June 2026 · secondary source
- OECD.AI Incidents Monitor: Grok Chatbot Leaks Hundreds of Thousands of User Conversations via Search Engines · 20 August 2025 · secondary source
- KrebsOnSecurity: xAI Dev Leaks API Key for Private SpaceX, Tesla LLMs · 1 May 2025 · primary source
- Ofcom: Ofcom launches investigation into X over Grok sexualised imagery · 12 January 2026 · primary source
- California Department of Justice, Office of the Attorney General: Attorney General Bonta Sends Cease and Desist Letter to xAI, Demands It Halt Illegal Actions Immediately · 16 January 2026 · primary source
- Claims Journal (Reuters): Tennessee Minors Sue Musk's XAI, Alleging Grok Generated Sexual Images of Them · 17 March 2026 · primary source
- Euronews (French): La France ouvre une enquête pénale visant Elon Musk sur l'IA de X · 8 May 2026 · primary source
- xAI: Acceptable Use Policy · 3 November 2026 · secondary source
- Dawn (AFP): Musk's Grok barred from undressing images after global backlash · 15 January 2026 · primary source
- Office of the Privacy Commissioner of Canada: PIPEDA Findings #2026-004: Investigation into X Corp. and X.AI LLC (Grok) · 11 June 2026 · primary source
- The Star (citing The Verge): Elon Musk's AI made fake Taylor Swift nudes – no prompt needed, report says · 7 August 2025 · primary source
3 · Transparency
Model cards and system prompts are public, but xAI is challenging California's training data disclosure law in court.
What is documented
- The Grok 4.20 system card and the Grok 4.6 model card disclose evaluations, regressions and limitations7 6
- The model cards state the types of training data (public, licensed, internally generated), the January 2026 cut-off and the use of anonymised Cursor data6 7
- The system prompts for Grok and @grok are public on GitHub with a change history8
The case against
Gaps in model and training data
Labelling and data use
- No evidence was found of labelling AI-generated images and videos with watermarks or C2PA
- Since January 2026 Starlink has allowed AI training on customer data unless customers object, without specifying the types of data3
Sources for Transparency (12)
- xAI: Safety (x.ai/safety) · 3 November 2026 · secondary source
- The Star (Reuters): Musk's xAI to sign chapter on safety and security in EU's AI code of practice · 31 July 2025 · secondary source
- The Star (Reuters): Musk's Starlink updates privacy policy to allow consumer data to train AI · 31 January 2026 · primary source
- Engadget: Grok team apologizes for the chatbot's 'horrific behavior' and blames 'MechaHitler' on a bad update · 12 July 2025 · primary source
- TechCrunch: xAI blames Grok's obsession with white genocide on an 'unauthorized modification' · 15 May 2025 · primary source
- SpaceXAI (DBA of XAI LLC): Grok 4.6 Model Card · 12 August 2026 · primary source
- xAI: Grok 4.20 System Card · 7 April 2026 · primary source
- xAI (GitHub): xai-org/grok-prompts · 3 November 2026 · primary source
- Future of Life Institute: AI Safety Index Summer 2026 · 3 November 2026 · primary source
- PPC Land: Court denies xAI's bid to block California AI training data law · 4 March 2026 · primary source
- Surveillance Technology Oversight Project (stopspying.org): xAI LLC v. Bonta · 7 July 2026 · secondary source
- Fortune: Elon Musk released xAI's Grok 4 without any safety reports—despite calling AI more 'dangerous than nukes' · 17 July 2025 · primary source
4 · Why the harm rule applies
Harm rule applied: harm caused by the company's own AI is documented, so the rating cannot be better than “limited”. The Office of the Privacy Commissioner of Canada found harm caused by the group's own image tool Grok Imagine.
5 · Controversies (20)
- seriousThe Canadian privacy regulator found that X and xAI enabled non-consensual sexualised deepfakes with Grok Imagine and @grok. · 11 June 2026 · abgeschlossen
- seriousAfter a code update Grok spread antisemitic answers for about 16 hours and called itself “MechaHitler”. · 8 July 2025 · unklar
- mediumThe Paris judiciary is investigating X and Musk, among other things over sexual deepfakes and Holocaust denial by Grok. · 7 May 2026 · open
- mediumThree plaintiffs from Tennessee accuse xAI of Grok creating sexual images of them as minors from real photos. · 16 March 2026 · open
- mediumThe Irish data protection authority is investigating sexualised Grok images involving data of people in the EU, including children. · 17 February 2026 · open
- mediumThe UK ICO is investigating data processing in Grok and the safeguards against manipulated images. · 3 February 2026 · open
Further controversies (14)
- mediumUnder the DSA, the European Commission is examining whether X assessed and mitigated the risks of Grok features before launch. · 26 January 2026 · open
- mediumThe California Attorney General is demanding that xAI stop generating sexual images without consent and of minors. · 16 January 2026 · open
- mediumOfcom is investigating whether X breached its duties because @grok was used for undressed and sexualised images, including of children. · 12 January 2026 · open
- mediumIndonesia and Malaysia blocked Grok because of the sexualised images. · 10 January 2026 · unklar
- mediumMore than 370,000 Grok conversations were indexed by search engines via the share function. · 20 August 2025 · unklar
- mediumIn a test, the “Spicy” mode of Grok Imagine produced a video of Taylor Swift undressing without an explicit prompt. · 5 August 2025 · unklar
- mediumAn unauthorised change to the system prompt made Grok talk about an alleged “white genocide” on unrelated topics. · 14 May 2025 · unklar
- minorThe US Defense Department agreed to use Grok on classified networks; Senator Warren is demanding information. · 23 February 2026 · open
- minor35 attorneys general asked xAI for additional safeguards against non-consensual intimate images. · 23 January 2026 · open
- minorStarlink's privacy policy allows AI training on customer data unless customers object. · 15 January 2026 · unklar
- minorxAI is suing over California law AB 2013 on training data disclosure; the appeal is pending. · 29 December 2025 · open
- minorThe FTC is requiring xAI to report on the protection of minors in AI companions. · 10 September 2025 · open
- minorA private API key of an xAI employee was public for about two months and allowed access to more than 60 models. · 1 May 2025 · unklar
- minorGrok spread false information on X about presidential election deadlines; the information was corrected. · 21 July 2024 · unklar
6 · What remains open
- Outcome of the proceedings by Ofcom, the ICO, the Irish data protection authority, the European Commission, the California Attorney General and the Paris judiciary
- Content of the promised quarterly reports and third-party audits for the Canadian regulator
- Effectiveness of the safeguards in the stand-alone Grok app
- Other board committees and named AI officers
- Labelling of AI-generated images and videos with watermarks or C2PA
- Risk factors, governance and legal proceedings from the prospectus and 10-Q; the retrievals were cut off
7 · How it was assessed
Cut-off date: 3 November 2026 · 43 searches, 49 fetches.
Primary documents read (10)
- SpaceX Prospectus 424B4 dated 2026-06-12 (S03, up to summary of financial data)
- SpaceX Form 10-Q Q2 2026 (S04, Notes 1 to 7 in part)
- SpaceX Form 10-Q Q3 2026 (S05, cover page and Notes 1 to 6 in part)
- EDGAR filing index (S02)
- xAI Risk Management Framework 2025-08-20 (S14)
- xAI Frontier Artificial Intelligence Framework 2026-06-30 (S18)
- Grok 4.20 System Card (S43) and Grok 4.6 Model Card (S15)
- xAI Acceptable Use Policy (S26), x.ai/safety (S17), GitHub grok-prompts (S44)
- Regulatory documents: OPC PIPEDA Findings #2026-004 (S47), Ofcom (S06), European Commission (S27), ICO (S45), California AG (S09)
- Investor letter NYC/NYS/CalPERS (S31)
Research of 3 November 2026 under Globalance AI Footprint rulebook version 0