Google has quietly moved one of Chrome’s most sensitive security checks off your machine. The browser’s AI-driven scam detection, which once ran entirely on-device, now sends page text to a server-side Gemini model for analysis — a shift the company frames as a horsepower upgrade rather than a privacy trade-off. The change was disclosed on October 20 by Jasika Bawa, group product manager for Chrome, alongside two other 2026 defences: a new warning tier for “suspicious” sites and clipboard monitoring aimed at command-based malware.
- Off-device analysis: Full page text now travels to Google’s servers, not just the condensed signals promised when the feature debuted in May 2025.
- An amber tier: Chrome adds low-confidence “suspicious site” warnings, designed so the browser can flag more pages before they are confirmed dangerous.
- Clipboard inspection: Chrome now scans commands copied from web pages, strips personal data and checks the remainder against Safe Browsing — a direct answer to ClickFix attacks.
- Opt-in only: All three features require Enhanced Protection, which Google says covers more than a billion Chrome users.
Why the on-device promise was dropped
When Google launched the capability in May 2025, local inference was the selling point: the model ran on the user’s hardware and only summarised signals reached Google’s infrastructure. That architecture has now been inverted, and the rationale is coverage. A server-side model can serve every Enhanced Protection user instead of only those with compatible devices.
Google attaches three limits. Text is transmitted only when a page already exhibits high-confidence scam indicators; processing is ephemeral; and the content is, in the company’s words:
never logged, retained or used to train models.
Even so, the substance of what leaves the device has changed. According to Google, Gemini weighs which brand a page appears to impersonate, what the page is trying to achieve and an AI-generated scam-likelihood score, before Safe Browsing merges those findings with domain reputation and other signals. The scope has widened as well, from fake tech-support pages to bogus antivirus alerts and counterfeit browser-update prompts.
An amber light before the red screen
Until now, Chrome’s flagship warning was the full-page red interstitial reserved for sites Google can prove are dangerous. That screen stays. The new “suspicious sites” tier covers pages showing deceptive patterns typical of known malicious sites, but where Google’s confidence is lower. The company says the alerts are distinguishable by format, iconography, colour and wording, and has not specified the exact colour.
The stated logic is revealing: by separating low-confidence warnings from high-confidence ones, Chrome can afford to show more of them. That carries a cost the announcement never quantifies — false positives — and it matters more as AI tooling makes cloned storefronts, banking portals and support pages harder to tell apart from the real thing.
Clipboard checks and the ClickFix economy
The third defence is the most invasive and the least documented. ClickFix attacks fake a technical fault and instruct victims to copy a command and run it, usually through PowerShell or the Windows Command Prompt. Because the payload executes in the operating system rather than in the browser, Chrome’s download scanning never sees it.
For Enhanced Protection users, Chrome now detects when a copied snippet resembles such a command, strips personally identifiable information, submits the remainder to Safe Browsing and raises a warning if the verdict is malicious — all before execution. Google says the filter was tuned for high precision so developers pasting legitimate commands are not interrupted. It has not published what counts as a suspicious command, what it treats as personal data, or how that separation is performed; clipboard inspection has never appeared in Chrome’s public documentation before.
Reading the “ten times safer” claim
All three features sit behind Enhanced Protection, an optional setting tucked into Chrome’s privacy and security controls. Google says it makes users ten times safer against phishing and social engineering than standard protection. Its own footnote explains the metric: it compares the average number of security warnings displayed to each group. Showing ten times more alerts is not the same as preventing ten times more harm — and the new suspicious-site tier exists precisely to display more alerts. No figure for scams actually blocked has been released.
The trade-off extends beyond the scam model. Enhanced Protection already sends Google the address of each site a user visits, plus small samples of page content, extension activity and system information; standard protection instead checks a disguised fragment of each URL through privacy servers. Anyone managing a parent’s laptop or a handful of office machines now faces a decision worth making deliberately, and Google has not specified which Chrome versions, platforms or countries receive the update, or when.
What the rebranding of browser security really means
The direction of travel is sound. Attackers increasingly persuade users to compromise their own machines, and a browser that reads intent rather than matching signatures is a genuine advance over blocklists. But the framing is backwards. The headline for users is not a new colour of warning — it is that Chrome’s threat model now depends on page text and clipboard contents flowing into Google’s cloud.
Server-side inference is a defensible answer to the hardware lottery that capped the 2025 rollout, and the ephemeral-processing safeguards are meaningful if honoured. It is still an architectural privacy decision that deserved its own paragraph, not a subordinate clause about model power. The next test is transparency: documentation of the clipboard classifier, real-world false-positive rates and a rollout schedule would turn a vague promise of safety into something verifiable.
Sources & References