An AI assistant with your bank password, and nobody's blinking
Issue #022 · 4 min read
An AI assistant with your bank password, and nobody's blinking
Instinct's access looks great in a demo. Plus what's actually shippable this week from GitHub, Vercel and Linear.
The big story
Early testers of a new AI assistant called Instinct say it's great.
It can act on your behalf, across your accounts, without you doing the clicking.
That's also the problem. Instinct's terms give it sweeping access to a user's accounts and the ability to act on their behalf, and some testers are uneasy about it even while they praise what it can do — that's TechCrunch.
This is the same tension every agent product is racing toward right now, just stated plainly for once.
An assistant that's actually useful needs to read your email, touch your calendar, maybe move money or send messages as you.
An assistant with that much reach is also a single login away from being a disaster, and "broad terms" is doing a lot of quiet work in that sentence.
Separately, GitHub's own security team just made a smaller but related point: they added a way to dismiss a code scanning alert as "mitigated" rather than fixed, for cases where a hole still exists in the code but something else — a firewall, a network rule — is standing in front of it, per the GitHub changelog.
Sensible feature. But it's also an admission that "the hole is still there, we just built a wall in front of it" is now a normal, trackable state for software to be in.
If you're a PM shipping anything agentic this quarter, the Instinct story is worth reading closely before your next design review.
The question is never "can the agent do this." It's "what happens the day someone convinces it to."
What shipped
Linear's coding agent now sets up, runs, and tests your code before handing it back, and the company cut its AI credit pricing at the same time. Per the changelog, coding sessions now adapt automatically to each codebase's configuration, which means fewer half-finished handoffs landing back on an engineer's desk. Worth a look if your team is drowning in agent output that still needs a human to finish the last 20%.
GitHub Copilot now runs inside Slack and Microsoft Teams, in public preview. You mention @GitHub in a channel and it starts a session everyone in the thread can watch and steer — see the Slack and Teams posts. This is aimed squarely at the moment an engineer says "I'll just paste this in Slack" — now the agent lives where the conversation already is, instead of a tab nobody checks.
Vercel dropped the price of GPT-5.6 Sol by half through its AI Gateway. No caveats in the changelog about why, just a lower price, live now. If you're already routing model calls through Vercel's gateway, this is free margin. If you're not, it's a small reminder that model pricing keeps sliding and your unit economics from six months ago are probably stale.
What I'd actually do this week
- Pull up whatever agent tool in your stack has "act on behalf of the user" permissions and actually read what it can touch. Not the marketing page — the terms.
- If you're evaluating Linear's coding agent, check the new pricing against last month's bill before you assume the old numbers still hold.
- If any of your product surfaces let an agent take an action a human used to approve, write down what "mitigated but not fixed" would look like for that flow. GitHub just made that a formal state. Yours probably has one too, even if nobody's named it.
Reply and tell me which agent in your stack has access you'd rather it didn't.
Tools mentioned
- Instinct
- Linear coding sessions
- GitHub Copilot in Slack
- GitHub Copilot in Teams
- GPT-5.6 Sol pricing on Vercel AI Gateway