GitHub goes down worldwide as AI agents show security cracks

There's a nice irony buried in yesterday's news cycle: the same day GitHub went down for three-plus hours, knocking out Copilot, code review, and build pipelines for developers worldwide, another story surfaced about a Copilot-assisted code flaw that let Wiz's autonomous red-teaming agent breach Snowflake's internal Jira. Two very different failure modes, but the same underlying lesson — we've built an enormous amount of engineering infrastructure on top of AI-assisted tools and centralized platforms, and we're only now discovering how brittle both dependencies can be.

The GitHub outage is the easy one to explain and the harder one to fix. When your code review, CI/CD, and AI pairing tool all live behind the same login page, a single infrastructure failure doesn't just annoy you — it stops entire companies from shipping. This isn't really an AI story, but it's a preview of what happens as agentic workflows get layered on top of the same centralized dependencies. If Copilot goes down alongside the repo, that's inconvenient. If an autonomous coding agent goes down mid-deployment because its orchestration layer sits on a platform that's having a bad Monday, that's a production incident with a different shape entirely.

The Snowflake breach is the more interesting story, because it shows the security implications of agentic AI aren't hypothetical anymore. Wiz didn't find a theoretical vulnerability — they used an autonomous red-teaming agent to actually exploit a flaw that originated in a pull request Copilot helped write, and that exploit chain led into Snowflake's internal systems. This matters because it inverts the usual framing of "AI agents as productivity tool" into "AI agents as attack surface and attack vector simultaneously." The code-generation assistant introduced the weakness; a different AI agent found and exploited it. I expect we'll see a lot more of this pattern — offense and defense both increasingly automated, with humans mostly reviewing outcomes after the fact rather than the process itself.

Which brings me to the theme that keeps surfacing across enterprise AI coverage this week: everyone's adopting agents, almost nobody's running them safely in production. The surveys on agentic AI adoption tell a familiar story — pilots everywhere, production deployments rare — and Oracle's pitch about "operationalizing" agents at scale is really an admission that the hard part was never building the agent, it's governing what it's allowed to touch. Startups like xpander are trying to solve this by giving enterprises ownership of their own agent infrastructure rather than renting it piecemeal from frameworks like LangSmith or CrewAI, and their $7.5 million seed round plus a strong GAIA benchmark score suggests investors think control layers are where the real value will accumulate, not the agents themselves.

By the way, none of this is an argument against building with agents — it's an argument for treating the plumbing around them with the same seriousness as the models. The GAIA benchmark and the Snowflake breach both happened in the same week for a reason: capability is outpacing the infrastructure meant to contain it, and until that gap closes, every impressive demo comes with an asterisk.