ChatGPT, Grok, and Claude all went down within 90 minutes
Three of the biggest consumer AI products broke on the same morning, and nobody could say why. Between 9:30 and 11AM ET, Grok, Claude, and ChatGPT all started throwing errors — a rare simultaneous failure across rival labs that share almost no infrastructure.
Grok went first at 9:30AM ET, Claude followed at 10:30, and ChatGPT at 11:00. OpenAI's status page reported "elevated errors across ChatGPT and Codex," with logins, file uploads, voice mode, search, deep research, and image generation all affected; the company said it had "applied a mitigation" and was "monitoring recovery" while tools remained at degraded performance. Anthropic's status page initially listed Mythos/Fable 5.1, Mythos/Fable 5, Opus 5, Opus 4.8, and Opus 4.6 as impacted before narrowing to Opus 4.8 and Opus 5 — technical staff member CJ Avilla attributed it on X to an "infrastructure issue," adding "we're working on it but don't have an ETA yet." Grok's app on iOS, Android, and the web returned "This model is overloaded right now," and xAI's site said it was "working on restoring service as quickly as possible."
Three separate incidents are the boring explanation, and probably the right one. Nothing in the reporting points to a shared cause: different clouds, different stacks, staggered start times, and three separate status pages describing three different failure modes. But the coincidence lands at a moment when the honest answer to "what happens if the model goes away" is increasingly uncomfortable — OpenAI is teasing Astra, its next model, while shipping a chatbot that stopped accepting prompts for hours. The clustering also exposes how thin the consumer layer is: there is no fallback, no offline mode, and no meaningful SLA for the free and Plus tiers carrying most of the traffic. Millions of people discovered at once that a tool they now treat as infrastructure is really just someone else's service, and the labs are under no obligation to explain what broke.
Flock Safety trained police to monitor protests from a dashboard, not the street. In a webinar reviewed by 404 Media, Flock employees walked officers through using the company's cameras and law enforcement databases to watch No Kings demonstrations — overlaying live video, traffic data, building floor plans, and a "response plan" for the event on a single Denver map. A former New Orleans officer now working at Flock explained that the system pulls together license-plate reads, body camera footage, and records queries into "full situational awareness," and described choosing ALPR camera placements as making herself "dangerous if and when I need it." The same webinar detailed how 7 terabytes of footage from the 2025 Bourbon Street attack flowed through FlockOS to DHS, federal fusion centers, and the FBI.
This is the surveillance story that AI keeps getting dragged into, and the company's own training material is the evidence. Flock publicly frames its network as a violent-crime tool and downplays its reach, but the webinars show a product sold explicitly as protest infrastructure — with the pitch that remote monitoring avoids cops on the ground "making it worse, getting attention." That framing inverts the usual civil liberties argument: the less visible the policing, the harder it is to document or contest. The Electronic Frontier Foundation documented the same pattern last year; what's new is the vendor saying the quiet part on a recorded call.
CrowdStrike shipped an identity provider built for AI agents instead of humans. Announced at Fal.Con, the Agentic Identity Provider issues each agent a cryptographically verifiable identity and brokers short-lived, task-scoped tokens rather than handing over standing credentials — every action traces back to the human or workload the agent is acting for. CrowdStrike's pitch is that service accounts, API keys, and workload identities were never designed for software that acts autonomously and spawns sub-agents; Continuous Identity general manager Scott Kriz put it bluntly: "you cannot continuously authorize an identity you were never able to establish." Chief security officer Jennifer Johnson's framing of the scale problem — "you're going to have 90 agents for every human" — is the number worth remembering. Alongside it came parallel SOC investigations that chase an attack across endpoint, identity, SaaS, cloud, and network at once, plus command-line blocking of malicious packages before install scripts run. We covered the company's move into security-specific models Tuesday — CrowdStrike ships security frontier models built on Falcon telemetry.
What to watch: whether OpenAI, Anthropic, and xAI publish postmortems that rule a shared dependency in or out — three simultaneous failures with no common cause is a coincidence worth documenting properly.
When three rival AI products vanish in the same hour, does that change how much you'd trust one with something you can't afford to lose? Tell us in the comments.
Sources: The Verge · MacRumors · Axios · 404 Media · EFF · SiliconANGLE · SiliconANGLE — CrowdStrike announces
