A hacker used AI agents to break into John Deere and SharkNinja
Issue #062 · 4 min read
A hacker used AI agents to break into John Deere and SharkNinja
Plus: GitHub kills off Copilot models overnight, and Microsoft says treat agents like insiders
The big story
One hacker. Two very different companies. Both broken into.
A hacker broke into John Deere and SharkNinja using an army of AI agents, not a crew of people — The Verge has the story.
That's the headline fact, and it's worth sitting with before you move on.
This isn't a hypothetical about what agents might someday do. It happened, to two real companies, this week.
It also lines up with something Microsoft said in the same news cycle: enterprises should treat AI agents like insider threats, because "frontier models can't be trusted" — that's from TechCrunch's coverage.
Microsoft's point is specifically about agents running on systems the big AI labs don't control. Put the two stories together and the picture is uncomfortable: the same agents your team might be piloting for productivity are, in someone else's hands, a break-in kit.
There's a third data point too. OpenCode, an open-source coding agent that's become a popular Claude Code and Cursor alternative, got hit by hackers who stole users' API keys — and the AI tools OpenCode turned to for help kept refusing to even look at the malicious code, according to TechCrunch.
The pattern across all three: agents are getting good enough to do real damage, and the defenses built around them aren't catching up at the same speed.
If you're a PM shipping anything with agent access to your systems, this is the week to ask your security team exactly what an agent can touch, and whether anyone is watching it the way they'd watch a new hire with a laptop.
What shipped
GitHub deprecated a batch of Copilot models on October 19, effective immediately, across Copilot Chat, inline edits, ask and agent modes, and code completions — see the changelog.
If your team has prompts, docs, or internal tooling that name a specific model, check them now, not after something breaks.
GitHub also posted a second notice: more Copilot models are scheduled to go away on November 16, listed here. Two deprecation waves in a month is a pace worth flagging to anyone who owns a Copilot rollout.
Alongside that, GitHub shipped a Copilot model policy API, letting enterprises and orgs retrieve and update which models their teams can use via REST calls instead of clicking through settings — details here. That's a direct answer to the deprecation churn above: if you're managing Copilot access for more than a handful of people, this is how you'll want to enforce policy going forward rather than chasing changelog posts.
Otter.ai raised $320 million for its meeting note-taking tool, from General Catalyst's Customer Value Fund — a structure that funds growth without taking equity and gets paid back out of the results of marketing spend, per TechCrunch.
Otter says the money goes to enterprise growth, go-to-market, international expansion, and product work. Worth watching if your org already has Otter seats: a funding round this size usually means a pricing or packaging change isn't far behind.
What I'd actually do this week
- Audit what your agents can actually touch. Not what they're supposed to do — what they're permitted to do if something goes wrong. The John Deere and SharkNinja break-in didn't need a new exploit, just agents with reach.
- Grep your docs and prompts for hardcoded Copilot model names. GitHub just killed one batch and scheduled another for November 16. Five minutes now beats a broken workflow later.
- If you manage Copilot for a team, look at the new model policy API and decide whether you want model choice centrally enforced instead of left to individual settings.
Reply if you've already had an agent do something in production you didn't expect — I'd like to hear it.