counter hit make
Trending

AI Tools

Google Says AI Attacks Crossed a Threshold. Ask Who's Holding the Ruler.

4 min read

The tool is rarely the bottleneck. The process around it is.
The tool is rarely the bottleneck. The process around it is.

The most useful thing you can do with Google's new warning about AI-driven cyberattacks is neither panic nor shrug. It's to notice the packaging: the alarm arrived as a company blog post, on a company site, from a company that sells the detection tools, the cloud infrastructure, and the AI models in question.

That doesn't make it false. It makes it unaudited.

Google's Threat Intelligence Group says agentic threat activity has crossed "a significant threshold," and the Q4 tracker is now circulating through wire coverage and security newsletters. The phrase is doing enormous work. A threshold sounds like a gauge. Gauges have units, baselines, and someone other than the manufacturer reading the dial. So before this becomes the sentence you repeat at a budget meeting, run it through the three failure modes that catch most scary-number stories.

First: who benefits from you believing it

Google Cloud sells security operations tooling. Google sells Gemini. Google sells the infrastructure a large share of the internet runs on. A report that says the threat has fundamentally changed, published by the vendor whose product category is the response, is not disqualifying — but it is a structural conflict you should price in.

This is not a Google problem. CrowdStrike, Microsoft, Palo Alto Networks and every other major vendor publish annual threat research that reliably discovers the exact threat their product was built to stop. Some of that research is genuinely excellent, and Google's threat group has a real track record on nation-state activity. But excellent research and self-interested framing coexist comfortably in the same document. The tell is never the finding. It's the recommendation section.

Second: what's missing

Counts without denominators. "More agentic attacks observed" is perfectly consistent with "more agents looking." If a vendor turns on AI-assisted detection across its telemetry, the tally of AI-assisted intrusions goes up even if the world outside the sensor doesn't change by a single incident. Epidemiologists have a name for this — surveillance bias — and it's the single most common way a security statistic gets inflated without anyone lying.

The questions a rigorous version of this claim would answer: How many incidents were actually investigated? In how many did an autonomous system take a consequential action — moving laterally, buying a domain, retrying a failed login — without a human steering it? Of those, how many succeeded? Against whom? And what's the comparison period?

Attribution is the other gap. Proving a machine ran an intrusion is hard. It usually rests on observing the attacker's own tooling or on stylistic fingerprints in the text — and fingerprints are precisely what defenders train classifiers on, which means the signal degrades the moment it becomes public. A confident label like "AI-driven" is a conclusion, not evidence, and the public write-ups rarely walk you through the chain.

Third: what would have to be true

For "agentic" to mean what a reader thinks it means, the system needs to plan, adapt and execute with minimal human direction: choose targets, acquire credentials, escalate, persist, and recover from its own mistakes. That is a real capability frontier, and if it has genuinely been crossed in the wild, it deserves the alarm.

But there's a weaker version that produces nearly identical headlines: a language model writes a more convincing phishing email, in flawless English, in five languages, at ten thousand messages an hour, and a human being still clicks send and still cashes the wire. That is dangerous. It is also a cost reduction for criminals, not a threshold — a continuation of a trend that has been running since spam filters got good. The two versions demand completely different responses, and "significant threshold" blurs them.

Here's the uncomfortable middle ground. Even the weak version matters more to a 40-person accounting firm in Ohio than to a Fortune 500 with an incident response retainer. Google's report is about a capability shift; the damage lands on organizations with no security team, no cyber insurance worth the premium, and a controller who takes vendor payment-change requests by email because that's how it's always been done. Business email compromise, W-2 phishing in tax season, wire fraud against real estate closings — that's where an AI that writes better English converts directly into dollars.

What actually survives scrutiny

Strip out the framing and the defensible core is this: the labor cost of a competent intrusion and the language cost of a competent con are both falling, and defenders should assume the volume and quality of targeted attempts rises regardless of whether any single attack was fully autonomous. You don't need a threshold to justify fixing the basics. You need only the direction of travel.

So watch for corroboration, not amplification. If the finding is real, independent incident response firms, insurers and CISA will describe the same pattern within a quarter or two, from their own case files, without citing Google. If only the vendor selling the countermeasure can see the threat, that tells you something about the sensor.

Meanwhile, do the boring work that defeats both the strong and weak versions. Move to phishing-resistant multi-factor authentication — hardware keys or passkeys, not six-digit codes. Require a callback to a known number before any change to payment instructions, no exceptions, no matter how urgent the email sounds. Keep one backup offline and test a restore. Patch internet-facing systems on a schedule you can name out loud. And ask your IT provider which of these they've actually verified, not which ones they've recommended.

The threshold claim may turn out to be right. Bet on the direction of travel anyway, because the cost of being wrong about the date is much lower than the cost of being wrong about the basics.

Common questions

Did Google say AI is now carrying out cyberattacks on its own?

Google's Threat Intelligence Group says agentic threat activity has crossed what it calls a significant threshold. The distinction that matters is how much human direction was involved in each case — an autonomous system choosing targets and moving through a network is a different claim from a model drafting convincing phishing text that a person then sends. Read the report for the specific incidents, not the framing.

Should I distrust vendor threat reports entirely?

No — distrust is the wrong setting. Treat them as credible but interested. Vendor research is often the best public window into attacker behavior, and it is also published by companies selling the remedy. Ask what the denominator is, how attribution was established, and whether anyone outside the vendor has seen the same pattern.

What should a small business do first, given limited budget?

Phishing-resistant multi-factor authentication, a mandatory callback to a known number before any change in payment instructions, one offline backup you have actually restored from, and a patch schedule for internet-facing systems. Those four close the paths most commonly used in wire fraud and ransomware, regardless of whether the attacker was a person or an agent.

Does this mean AI is now more dangerous than human attackers?

No. The evidence points to AI lowering the cost and raising the quality of attacks that humans still run and still profit from. That is serious, and it is not the same as machines operating a campaign end to end.

Related guides