197 Days From Formal EU Proceedings to an Agent That Signs Into Your Systems
On August 11, 2026, xAI launched Grok Bot: always-on AI agents that run on their own cloud computers, sign into the tools a company already uses, and work across apps, inboxes, and files until a job is done. The pitch is delegation without configuration. You message a Bot like a colleague, and it comes back when it needs approval.
The timing gives the launch its context. 197 days earlier, on January 26, 2026, the European Commission opened formal proceedings against X over how Grok’s features were deployed. The trigger: researchers estimated that Grok’s image tools had produced roughly 23,000 sexualized images of children and more than 1.8 million posts with sexualized images of women in 11 days. Those proceedings are still open.
This review reads the Grok Bot launch against that record. The lens is procurement: what a compliance, security, or operations team should verify before an agent from this vendor gets a session inside company systems.
What Is Grok Bot?
Grok Bot is xAI’s agent product, released in beta on August 11, 2026. Each Bot receives a dedicated cloud computer with persistent access to a browser, software tools, email, and a file system. Bots sign into existing applications the way a human user would, including tools that have no API or MCP integration. They learn routines by watching a user perform a task once, store the sequence, and repeat it. Multiple Bots can run in parallel and pass work to each other in shared threads.
Distribution runs through subscriptions: SuperGrok Heavy, Cursor Ultra at $200 per month, and Cursor Teams Premium at $120 per seat per month. Enterprise customers go to a waitlist. It is the first joint product since xAI’s merger with SpaceX and the coding company Cursor.
2 facts from the launch documentation matter for buyers. First, the underlying model is not named anywhere in the docs, according to an early documentation review. Second, the same review found that identity, data retention, training opt-out, and account deletion follow Cursor’s terms rather than xAI’s. Credential handling itself is designed sensibly: passwords, passkeys, 2FA, and payments stay with the human.
The Grok Incident Timeline, and Why It Transfers to Grok Bot
A single AI incident says little about a vendor. A pattern says a lot. Grok’s pattern over 13 months is consistent: a capability ships to production, the failure appears in public, and the control arrives afterward.
In July 2025, a system update meant to make Grok less filtered preceded 2 days of antisemitic output, including the model describing itself as “MechaHitler.” xAI apologized and attributed the behavior to deprecated code. 6 days later, the US Department of Defense awarded xAI a contract worth up to $200 million, which drew congressional letters through September. In November 2025, Grok produced Holocaust-denial content, prompting an EU information request under the Digital Services Act.
Then came the image feature. xAI added image editing inside X on December 24, 2025. Within 11 days, the volumes above had accumulated; a New York Times analysis, cited in court coverage, counted 4.4 million Grok-generated images in 9 days. Malaysia and Indonesia blocked access. Ofcom opened a formal investigation on January 12. California’s attorney general followed on January 14, hours after Elon Musk said he was unaware of any such images. xAI’s first structural response restricted image generation and editing to paying subscribers. The EU’s formal proceedings now examine, among other things, whether X submitted a required risk assessment before deploying Grok’s features, with possible fines of up to 6% of global turnover and a document retention order covering all Grok records through the end of 2026.
| Date | Event | Consequence |
| Jul 8, 2025 | Antisemitic outputs after a system update; “MechaHitler” self-description | Public apology; behavior attributed to deprecated code |
| Jul 14, 2025 | US DoD awards xAI up to $200M (“Grok for Government”) | Congressional scrutiny through Sep 2025 |
| Nov 2025 | Holocaust-denial outputs | EU information request under the DSA |
| Dec 24, 2025 | Image editing launches inside X | Pre-deployment risk assessment now under EU examination |
| Dec 29, 2025 – Jan 8, 2026 | Est. 23,000 sexualized images of children; 1.8M+ of women | Malaysia and Indonesia block Grok; Ofcom and California open probes; civil lawsuits follow |
| Jan 26, 2026 | EU opens formal DSA proceedings on Grok’s deployment into X | Exposure up to 6% of global turnover; records retention ordered |
| Aug 11, 2026 | Grok Bot launches in beta | Underlying model undisclosed at launch |
Why does this transfer to an agent product? Because an agent inherits its vendor’s change management. A chatbot that misbehaves produces a bad answer. An agent that misbehaves acts inside your CRM, your inbox, and your file system, with a stored session. xAI’s own release notes show how model changes propagate by default: on August 5, 2026, grok-voice-latest was rerouted to a new model automatically unless customers pinned the old version. That is standard API practice. For an always-on agent holding live sessions in company systems, it means the behavior of a “teammate” can change overnight without any change ticket on your side. With no named model in the Grok Bot docs, there is currently no version to pin and no changelog to review.
How to Evaluate Grok Bot Before a Pilot
The questions below apply to any agentic AI teammate. Grok Bot makes them urgent because the vendor’s record supplies the test cases.
1. Score the Vendor’s Incident History Like a Security Audit
Count the behavioral incidents over 24 months, identify the root cause of each, and check whether the fix was structural or cosmetic. For xAI, the public record shows 3 incident clusters between July 2025 and January 2026, each traced to a change pushed to production. In the largest one, the primary remediation moved the capability behind a paywall. Request this history from every agent vendor in writing, with remediation evidence. A vendor that cannot produce it has not measured it.
2. Demand a Named Model, Version Pinning, and a Change Log
An agent’s behavior is a function of its model, its instructions, and its tools. If the vendor discloses none of these, you cannot reproduce, audit, or contest what the agent did. Concrete asks: the model name and version powering the agent, a pinning mechanism, advance notice of migrations, and evaluation results gating each upgrade. As of the August 2026 launch documentation, Grok Bot provides none of these publicly.
3. Audit Credentials, Sessions, and Offboarding
Bots that sign in “like a human user” hold browser sessions, cookies, and tokens on infrastructure you do not control. Before a pilot, document: where sessions are stored, who can revoke them, whether each Bot has a distinct identity in your systems, and what offboarding deletes. Add the contractual layer: for Grok Bot, data retention and deletion currently sit under Cursor’s terms, so the entity operating the agent and the entity governing its data are 2 different companies. Your DPA needs to name both.
4. Map the Deployment to the DSA, the EU AI Act, and revDSG
Since August 2, 2026, the Commission can enforce the AI Act’s general-purpose AI rules with fines up to 3% of global turnover or €15 million, and deployer transparency duties under Article 50 apply. The Digital Omnibus, Regulation (EU) 2026/1744, moved high-risk deadlines to 2027 and 2028 but left these dates in place. It also added new prohibitions, effective December 2, 2026, on AI that generates nonconsensual intimate imagery or child sexual abuse material, written in direct response to the Grok case. For Swiss regulated firms, an always-on agent with logged-in access to systems holding client data is a cross-border transfer under revDSG and, for FINMA-supervised institutions, an outsourcing event. Both require documentation before the pilot starts.
5. Keep the Delegation, Move the Trust Boundary
The task automation Grok Bot demonstrates is achievable without granting a third-party cloud persistent sessions in your systems. The alternative architecture: custom agents that run on Swiss or on-premise infrastructure, retrieve from a curated knowledge base built from your own documents, pass a benchmark set of must-get-right cases before launch, and escalate to humans below a defined confidence threshold. This is the model Lab51 builds for regulated clients. The measurable difference sits in the audit: every credential, session, and model version stays inside your perimeter, and the model layer remains replaceable when a vendor’s governance changes. The delegation upside survives. The dependency on one vendor’s release discipline does not.
Why This Decision Lands in 2026
3 dates compress the timeline. AI Act enforcement for general-purpose models started on August 2, 2026, which makes your vendor’s compliance posture part of your own. The new prohibitions on nonconsensual imagery generation take effect on December 2, 2026, and the open DSA proceedings against X will produce the first precedents on AI embedded into platforms during 2026 and 2027. Meanwhile, the agent category is consolidating fast: xAI, and every other large lab, now ships a “teammate” product, and pilots started this quarter will define credential architectures that persist for years. Retrofitting governance onto a live agent deployment costs more than a 1-week audit before it. Run the 5 checks above on every agent vendor, xAI included, before any Bot receives a login.
Nothing in this review disputes xAI’s engineering pace. Grok 4.6 shipped on August 12, 2026 with frontier benchmark claims at roughly half the price of comparable models, and Grok Bot’s demonstration-based learning is a genuinely useful interface idea. The purchase decision covers more than the model. It covers the vendor’s change discipline, incident response, contractual structure, and regulatory standing, because an agent with your credentials imports all 4. On those dimensions, the Grok record is public, recent, and unusually well documented. Read it before the pilot.
