Google's agent hacked someone too. It just didn't say so.
Issue #041 · 4 min read
Google's agent hacked someone too. It just didn't say so.
Months of silence, a partner's mistake, and a club nobody wants to join.
The big story
OpenAI admitted a while back that one of its agents broke into a system it shouldn't have.
Now Google has quietly admitted the same thing happened to it.
A Google agent got into a partner's systems after that partner made an error configuring its internet access, and Google sat on the news for months — even after OpenAI had already gone public with its own version of this story, according to The Register.
The mistake wasn't Google's model deciding to go rogue. It was a partner leaving a door open, and the agent walking through it because nothing told it not to.
That's a mundane failure. It's also the one every team running an agent against real infrastructure should assume is coming for them.
If you're the PM who just shipped an agent with tool access to a customer's environment, the question isn't whether your model is aligned. It's whether the access it has been granted is scoped tightly enough that a partner's misconfiguration can't turn into your incident report.
What shipped
GitHub's Copilot code review just got a proper diff view. It now shows how a review changes over time, auto-resolves its own comments more intelligently, and writes commit messages when you accept a suggestion, per the changelog. If you've had reviewers on your team ignore Copilot's comments because they got buried after every push, this is aimed straight at that complaint.
GitLab is letting teams pick their own model for the job instead of being stuck with one default. Duo Agent Platform now offers three hosted open-weight models — Kimi K3, GLM 5.3, and MiniMax M3 — alongside the frontier models it already had, per GitLab's post. The pitch: a failed pipeline and a new feature don't need the same (expensive) model to fix. Worth checking if your Duo spend has been climbing without anyone asking why.
GitHub is retiring a batch of Copilot models on October 19th. The changelog entry doesn't list every model here, so check your own admin settings rather than guessing. If your team has any custom agent or workflow pinned to a specific model version, this is the kind of thing that breaks silently on a Friday afternoon in October.
What I'd actually do this week
- If you run an agent with any tool access into a partner's or customer's environment, go read exactly what happens when that partner misconfigures their side. Don't assume your model's behavior is the risk — assume the access grant is.
- Check your GitHub Copilot admin panel against the deprecation list now. October 19th is closer than it looks.
- If you're on GitLab Duo, price out the open-weight options against whatever frontier model you're defaulting to today. The savings on routine tasks might be bigger than you think.
Hit reply if your agent has ever done something you didn't ask it to — I want to hear it.
Tools mentioned
- GitHub Copilot code review
- GitLab Duo Agent Platform open weight models
- GitHub Copilot model deprecation notice