In Opposite Directions
Google turned on an agent that never sleeps. Anthropic turned down a guardrail. Same week. Same industry. Two very different signals about who is actually in charge.
Let’s start with Google.
They released a new model called Gemini 3.7 Flash. It powers something called Spark. Spark is not a chatbot you talk to when you feel like it. Spark runs all day and all night. It works inside your email, your files, your documents. It does not wait for you to ask.
Google calls that convenience. And for a lot of people, it will feel like convenience. Files get organized. Emails get drafted. Status reports get updated. All while you are doing something else entirely.
But here is the question nobody answers in the press release: who is watching Spark while it works?
Not a person, checking every action. Spark is built to act without you reviewing each step. That is the whole point of it. The oversight has to live somewhere else — in the design, in the limits, in what the agent is and is not allowed to touch.
Right now, most companies do not have that in place. Two out of three enterprises have no agent governance policy at all. Not weak policy. No policy. Spark is landing in a world that is not ready to govern it.
That is not a reason to panic. It is a reason to notice. An agent with standing permission to act, running continuously, inside your work — that is a different animal than a model you type a question into. It needs a different kind of leash. Most organizations have not built one yet.
Now here is the other half of the week.
Eight days ago, researchers at Stanford and the Arc Institute published something remarkable. They showed that AI models can design fully functional viral genomes. Not describe one. Design one, from scratch.
That same day — the very same day — Anthropic loosened the safety controls on their model, Fable 5, for biology questions. They cut the number of blocked benign questions by 85 percent. Doctors, patients, students — people asking normal biology questions — used to get shut down and rerouted to a weaker model. Now they mostly don’t.
Anthropic says the guardrails that matter most are still standing. Virology, toxicology, anything that looks like weapons-grade research — that still gets blocked. They are not opening the whole door. Just the part that was catching innocent questions by mistake.
Maybe that is the right call. It probably is. But look at the timing. The same week the world found out AI can design viruses, one of the most safety-focused labs in the business loosened its own biology guardrails. Not tightened. Loosened.
That is not a scandal. It is a pattern: safety controls are not fixed. They move. Sometimes they move because the company found a real improvement. Sometimes they move because pressure built up and something had to give. From the outside, you cannot always tell which.
That is the real lesson buried in both of these stories. Vendor safety promises are not a wall. They are a setting. Settings get adjusted. An agent’s permissions get expanded. A classifier gets relaxed. A safeguard gets updated in a blog post nobody outside compliance will ever read.
If your whole plan is “trust what the vendor says today,” you have no plan. You have a snapshot. And snapshots go stale.
This is exactly why chosen conduct has to sit alongside vendor controls, not underneath them. A company’s safety settings are theirs to change. What is not theirs to change is whether the humans and systems relying on that model are actually watching for the change when it happens.
Spark needs task limits and a kill switch that does not depend on Google’s mercy. Fable’s relaxed guardrails need someone tracking what changed and why, not just accepting the new default because it arrived quietly in a blog post.
Two companies. Two different moves. Same underlying truth. The rules can shift under your feet, in either direction, and usually without much warning. The only real defense is a standard that does not depend on the vendor holding still — one built to notice the shift and hold its own line regardless of which way the company moves.
This post was drafted with AI governed assistance and reviewed and directed by Michael S. Faust Sr. before publication.
Contact: micvicfaust@gmail.com
© 2026 The Faust Baseline LLC | All Rights Reserved
