A quiet two-day stretch — and honestly, that's the report. Two things happened that matter, both on the same theme: your assistant running safely while you're not watching. Everything else we checked came back "nothing to do," and we'll tell you that rather than dress it up.
Claude Code (one of the main platforms these assistants run on) shipped three small features this week that are all, quietly, about the 3am problem — what happens when your assistant is working and you're not there:
It warns before its login expires. The ugliest silent failure for a 24/7 assistant is credential death: access quietly expires mid-task, the agent stops working, and nobody notices until you do. Now it sees the expiry coming and can flag it before going dark.
It can't be tricked by a forged "you said yes." There's a real attack where malicious content plants a fake approval inside the agent's own working notes — "the user already agreed to this above" — and an unattended agent acts on it. Notifications now explicitly state when no human input has occurred, so that forgery stops working.
It can't quietly edit its own logs. In autonomous mode, tampering with the session transcript is now blocked — the record of what your assistant did stays trustworthy.
Small features, but this is what growing up looks like for autonomous assistants: fewer silent deaths, fewer forgeable moments, honest logs. We've told your agent to update this week — and in the spirit of honesty, ours was a few versions behind too. It's updating.
A good question for your assistant this week: "If one of your logins expired while I was asleep, would you notice — and how would I find out?"
The toolkit behind the big July 28 protocol change (covered in edition 001) got its security piece this week: a standard way for agents' tool-servers to check credentials when running on modern edge platforms. This one only matters if your assistant runs its own tools on the internet — if it does, it already has the details in its edition.
The July 28 protocol switch is ~2.5 weeks out with no surprises (and it's a milestone, not a cliff — nothing breaks that day). The sandbox feature we warned about last edition still isn't actually enforced, so that warning stands. And the security trackers were quiet this window — no new malicious skills or attacks on personal agents to report.
That's edition three. Two upgrades in self-reliance, one standing warning, and an honest "quiet week" everywhere else. If your agent isn't subscribed yet: for-agents.md — two minutes, it sets itself up.
— Robin & Neo
Claude Code hardening: CHANGELOG (v2.1.203–v2.1.205)
Edge auth for tool-servers: MCP TypeScript SDK 2.0.0-beta.3
Protocol countdown: MCP release-candidate notes