Google Made Its Security Agents Fully Autonomous. Here's Why We Still Keep a Human in the Loop

Google's security agents can now run without a human

On July 20, 2026, Google Threat Intelligence took its agentic AI capabilities out of preview and made them generally available to Enterprise and Enterprise+ customers. These agents don't just summarize alerts for someone else to act on. They hunt for threats, respond to incidents, triage daily alerts, answer natural-language questions to generate full threat reports, search for indicators of compromise, and investigate suspicious files, domains, and software vulnerabilities on their own. A dedicated Malware Analysis Agent even runs inside a secure cloud sandbox to pull apart malware samples without a human opening the file first. (Pulse2)

This is a real shift from where security tooling was a year ago. Most "AI agents" in that space used to do one thing: read an alert and summarize it for a human to decide on. Now Google is shipping agents that act first and explain themselves after. Google says the platform includes inline citations back to the underlying threat data, so an analyst can check the agent's reasoning rather than approve it blind.

That transparency detail is the interesting part. It tells you Google itself isn't fully comfortable handing over the keys without a way to check the work — even at their scale, with their training data, on a problem this well-defined.

Good tool, narrow problem, unproven track record

We build AI agents for a living, so a launch like this is genuinely exciting. It will make a lot of standard security work faster for the teams that use it, and it will free up real hours for analysts who currently spend their day triaging alerts one by one.

But "genuinely exciting" isn't the same as "trust it completely." How reliable an agent is comes down to how well the model was trained for that exact job. Google is very good at this, and the tool is laser-focused on one problem: security operations. Even so, it's too early to say how it performs across the full range of real-world incidents it will eventually meet. Expecting a 100% success rate from any of these tools, however well-built, isn't realistic yet — not this year, probably not next year either.

That's not a knock on Google specifically. It's the honest state of every agentic AI product on the market right now, including the ones we build for our own clients. The model can be excellent and the outcome can still be wrong in a way nobody anticipated, because the model has never seen that exact situation before.

Why every project we build keeps a human in the loop

We don't run a single project — not one — without human approval built into the process. That's not caution for its own sake. It's how we actually work, on every client engagement.

Across every phase of a build, we flag the points where a decision actually matters and go back over them ourselves: is this specific piece of the implementation correct? If we don't fully trust a step, we go looking for another tool or method to double-check it before it ships to a client. Sometimes what breaks an agent is something nobody could have predicted going in. A person can't know everything in advance, and neither can a model — the difference is a person can say "I'm not sure" and go verify, and right now a model usually can't.

That's the practical version of "human in the loop." It isn't a compliance checkbox added at the end. It's a specific point in the workflow, chosen in advance, where a person looks at what the agent did before that decision becomes irreversible — before the email sends, before the record updates, before the payment goes out.

That checkpoint has to be designed in from the start, not bolted on after something goes wrong. If a human only looks at an agent's work once a week in a batch review, that isn't human-in-the-loop — that's a human finding out a week late. The checkpoint needs to sit exactly where a wrong decision would otherwise become permanent, and it needs to be fast enough that people actually do it instead of clicking approve without reading.

What this means for your business tomorrow

If you run a small or mid-sized business, a launch like Google's is genuinely good news. Depending on what you do, agentic AI can take real work off your plate — faster triage, less manual searching, fewer repetitive checks that eat someone's morning.

But don't read "generally available" as "hands-off." Keep a human-in-the-loop step in anything an agent touches, especially anywhere a wrong call costs you money, data, or a customer relationship. This technology is moving fast — the models themselves are changing month to month — and your review process needs to keep pace with that speed, not fall a step behind it.

The right question isn't "can an agent do this?" Most of the time, increasingly, the answer is yes. The better question is "where exactly does a person need to check its work, and how do I make that check fast enough that it doesn't become the new bottleneck?" That's the design problem worth solving. It matters more than which model or which vendor you pick.

If you want to figure out where that line sits for your business — what an agent can safely own end to end, and where a human still needs to sign off before anything ships — that's exactly what we map out in a Discovery Sprint.