Grok generated sexualised images of children every 41 seconds. 23,000 in nine days. Every other AI company refused the same prompts. Baltimore sued. France raided. Three countries blocked it. Elon Musk called content restrictions "censorship." The uncensored AI generated child sexual abuse material at industrial scale. Grok Build had a privacy toggle that did nothing. The CLI silently uploaded your entire codebase — every file, every secret, every SSH key, the full git history — to a Google Cloud bucket. One test showed 5.1 GB uploaded when the task needed 192 KB. That's 27,800 times more data than necessary. Turning off "Improve the model" didn't stop it. Musk promised to delete the data. xAI never confirmed they did.
What they claim: xAI states Grok "is designed to be maximally helpful" and positions it as a truth-seeking AI assistant.
What we found: Grok is trained on X/Twitter data — a platform where misinformation, bot networks, and engagement-optimised content are endemic. Studies have found X has the highest rate of misinformation of any major social platform since content moderation was gutted in 2023. Training an AI on a platform optimised for engagement rather than accuracy, then calling it truth-seeking, is a fundamental contradiction. Grok has been documented generating fabricated citations, false statistics, and confidently wrong answers — the same hallucination problems as other LLMs but amplified by training data quality.
What they claim: Grok is marketed as having a "rebellious streak" and being willing to answer questions other AI models refuse.
What we found: Grok's "anti-woke" positioning led to documented instances of generating harmful content that other models refuse — including detailed instructions for dangerous activities, false medical claims, and politically inflammatory responses presented as fact. Researchers found Grok more willing to generate disinformation and less likely to add safety caveats. The "rebellious" marketing is a feature, but the consequence is an AI model that trades safety for engagement — the same dynamic that destroyed X's content moderation.
What they claim: xAI privacy policy states user data is used to "improve and develop our products."
What we found: Grok is integrated into X, which was acquired by xAI in March 2025. SpaceX then acquired xAI in February 2026. This means Grok training data, X user data, Starlink subscriber data, and Tesla vehicle data all flow to the same parent entity. Grok Voice Think Fast 1.0 handles Starlink sales calls, collecting names, addresses, phone numbers, email addresses, and account numbers. The same AI processes your social media posts, handles your phone calls, and powers ad targeting on X. One model, multiple data streams, no structural separation.
What they claim: xAI states it processes data in accordance with applicable privacy laws.
What we found: In September 2023, X updated its privacy policy to allow user posts to be used for AI training — retroactively applied to all existing content without individual consent. The EU Data Protection Board flagged this as a potential GDPR violation. Ireland's DPC opened an investigation. X offered an opt-out buried in settings, but the default was opt-in, and the training had already begun before the policy change was announced. Grok was trained on data that users posted under a previous privacy policy that did not include AI training.
What they claim: xAI positions itself as an independent AI company pursuing "understanding the universe."
What we found: xAI was acquired by SpaceX in February 2026. xAI previously acquired X in March 2025. Elon Musk controls Tesla, SpaceX, xAI, X, Starlink, and The Boring Company. xAI burns approximately $1 billion per month, funded by SpaceX profits ($8B annual). The company's independence is financial fiction — it survives on Starlink subscriber revenue and SpaceX government contracts. Musk personally directed X user data to train Grok, moved Nvidia GPUs originally allocated to Tesla to xAI, and merged the companies under SpaceX. Independence requires the ability to say no to your funder. xAI's funder is its owner.
What they claim: xAI promotes Grok as a truthful, uncensored AI assistant
What we found: After Elon Musk enabled image generation, Grok produced 4.4 million images in 11 days. CCDH estimated 23,000 were sexualised depictions of children — one every 41 seconds. Reuters tested: Grok produced sexualised imagery for 82% of prompts that OpenAI, Google, and Meta refused entirely. Baltimore became the first US city to sue xAI. France raided X's Paris offices. Malaysia, Indonesia, and the Philippines blocked Grok.
What they claim: xAI states it applies appropriate technical and organisational measures to protect user data, and runs a HackerOne programme so security researchers can responsibly disclose flaws.
What we found: On 3 June 2026 Adversa AI disclosed a cryptographic context-injection attack to xAI via HackerOne. Encrypted instructions hidden on an ordinary webpage are decrypted inside Grok's own Python runtime when a user asks it to summarise the page. The decrypted payload then directs the agent to build a fake "decryption key" containing the user's name, coarse location, subscription tier and the full set of prompts in the conversation, and to embed that string in URLs the agent loads by itself — transmitting it to attacker-controlled servers. Adversa describe it as "a true zero-click outcome triggered by an ordinary 'summarize this page' request", with "no user confirmation and no visible warning." The researchers chased xAI again on 4 and 10 August. On 19 August 2026 they confirmed the attack still reproduced — 77 days after disclosure, unpatched.
What they claim: Grok Build CLI has a "improve the model" toggle implying user control over data sharing
What we found: Security researcher cereblab discovered Grok Build CLI v0.2.93 silently uploaded entire Git repositories — including .env files with API keys, SSH keys, and full commit histories — to a Google Cloud Storage bucket (grok-code-session-traces). In one test, 5.1 GiB was uploaded when only 192 KB was needed (27,800x excess). Toggling "Improve the model" off did not prevent uploads. xAI issued no formal disclosure about scope of affected users or data retention.