Brave began a phased rollout on August 13, 2026 of protections that de-identify the graphics hardware data websites can read through the WebGL and WebGPU interfaces, removing one of the highest-entropy signals available to tracking scripts. The changes ship in version 1.93 and are enabled by default on desktop and Android.

Brave published the change as the 38th entry in its running series of privacy engineering posts. According to Brave, the work was carried out by Rohit Agarwal, a senior software engineer, and Peter Snyder, a principal researcher, and written up by Shivan Kaul Sahib, vice president of privacy and security. The post describes a rollout staged over several days rather than a single flip, meaning users on the same version may see different behaviour for a short period.

The target is narrow and technical. Two browser interfaces that exist to draw hardware-accelerated graphics also disclose, as a side effect, an unusually precise description of the machine underneath the browser. That description does not change when cookies are cleared, and it does not change when a user switches to private browsing.

What the graphics interfaces give away

WebGL and WebGPU let websites render graphics using the device's graphics processor rather than its main processor. According to Brave, the same interfaces expose the exact vendor and model of the graphics card, the graphics driver, and the precise set of features the hardware supports.

The post gives a concrete example of the debug string a WebGL query can return on an Apple laptop with an M-series chip, tested through the Electronic Frontier Foundation's Cover Your Tracks tool: 'ANGLE (Apple, ANGLE Metal Renderer: Apple M5 Max)'. The provenance of that string is itself notable. According to Brave, the debug extension that surfaces it was originally exposed for Google Maps by Chrome before being opened to all websites.

The vendor and renderer strings are not the only leak. The same WebGL context reports its full list of supported extensions, a set that varies with the combination of graphics processor and driver. That list can be hashed into a compact identifier. The newer WebGPU interface publishes its own hardware descriptors, returning an adapter's vendor, architecture and device fields, with the post citing a returned object of the form {vendor: 'apple', architecture: 'metal-3'}.

What makes these signals valuable to trackers is their stability. Graphics hardware does not change from one day to the next. A hashed combination of driver, model and extension list therefore survives cookie deletion, storage clearing and session boundaries, which is the property that distinguishes fingerprinting from state-based tracking. The technique combines many small device-specific signals into a single identifier that follows a user across sites without cookies and without consent.

Brave also reports an empirical finding from its own measurement work. The company ran a small web crawl of top websites and analysed the stack trace immediately preceding each call to the graphics interfaces. "We found that most websites were using these APIs exclusively for browser fingerprinting," the post states. Brave did not publish the sample size, the ranking source used to define top websites, or the date of the crawl, so the result cannot be independently replicated from the disclosure alone.

Three changes, two different strategies

The protections consist of three distinct interventions, and they do not all work the same way.

First, Brave replaces the WebGL vendor and renderer strings with a single generic string, so that every Brave user reports identical values. Second, it empties the WebGPU adapter descriptors. Third, it injects randomization into the WebGL extension list, so that hash-based fingerprinters see a different value per session, per site at the eTLD+1 level, and per storage area.

The split matters. The first two measures pursue uniformity: if all users of a browser report the same value, that value carries no information and the signal collapses to zero entropy. The third pursues randomization: the extension list is too functionally significant to blank out, so instead it is perturbed in a way that prevents a hashed value from being correlated across sites or across sessions.

Randomization and uniformity produce different failure modes for anyone on the receiving end. A blanked value looks like an absence. A randomized value looks like a series of contradictory presences. Systems that score inconsistency as a risk signal, rather than treating it as missing data, will read those two outcomes differently.

According to Brave, the objective across all three is the same. The company describes the goal as "to deny trackers a stable identifier while leaving websites everything they need to render rich experiences for their users."

Rollout, breakage and user controls

The protections were tested on Brave's Nightly and Beta channels over several months before the default rollout. Brave states it is optimistic the approach reduces the chance of breakage on most websites while still countering fingerprinting, and that it retains the ability to adjust the protections on a per-site basis where a site genuinely breaks.

Users are given three levels of escape. On any individual site they can switch off the graphics protections, disable fingerprinting protection entirely, or turn off Shields, the browser's blocking layer, altogether. Brave points to Cover Your Tracks as the verification route, noting that once the feature is fully deployed, the reported vendor and renderer values collapse to a generic string.

The post names desktop and Android. It does not mention iOS, where Brave, like every browser distributed through Apple's App Store, builds on the system rendering engine rather than its own Chromium fork.

Further work is flagged but not delivered. Brave says it plans to randomize the list of supported extensions exposed by WebGPU, which would extend the third intervention to the newer interface. No date is attached to that plan.

Brave also uses the post to draw a competitive line, arguing that most other browsers concerned with privacy fall into one of two categories: opt-in protections hidden behind a mode or a feature flag for expert users, or approaches that sacrifice usability by breaking functionality. "Privacy is for everyone, not just the most technical users," the post states. That framing is the company's own characterisation of its competitors, and the post offers no comparative measurement to support it.

Notably absent is a headline metric. Brave publishes no figure for how much identifiability the change removes.

A browser market splitting in two directions

Brave's move lands in the middle of a divergence that has been widening for eighteen months.

On one side, browser vendors have been tightening. Firefox 145 completed a second phase of anti-fingerprinting defences in November 2025, with Mozilla reporting that user trackability fell by close to half. Apple made Advanced Fingerprinting Protection the default for all browsing sessions in Safari 26, which restricts high-entropy interfaces including WebGL graphics information for scripts classified as fingerprinting tools. Mozilla later patched an IndexedDB flaw that had allowed a stable hidden identifier to survive private browsing and Tor identity resets, a reminder that the attack surface extends well past graphics.

On the other side, the largest advertising platform moved the opposite way. Google announced on December 18, 2024 that it would permit advertisers to use device fingerprinting and IP addresses, a change that took effect on February 16, 2025and drew a public rebuke from the UK Information Commissioner's Office, whose executive director of regulatory risk called the approach irresponsible. Chrome subsequently retired most Privacy Sandbox technologies in October 2025 after low adoption, having already abandoned third-party cookie deprecation in April 2025. Chrome did ship IP address masking for Incognito mode, though IP Protection was among the components later wound down.

The result is a web where the identification technique a browser engineer treats as an attack is, on the same calendar, a permitted targeting input in the largest programmatic platform's policy documents.

What it means for measurement and fraud systems

The direct commercial exposure runs through two channels rather than one.

The first is audience identification. Device graphs and probabilistic identity vendors treat graphics hardware descriptors as high-value entropy precisely because they persist. Removing the vendor and renderer strings on Brave traffic strips that entropy from any such system operating on those sessions.

The second is fraud and bot detection, and here the effect is less obvious. Anti-fraud vendors also read WebGL and WebGPU descriptors, using hardware and driver consistency as a heuristic for distinguishing real devices from automation. A session that reports a generic vendor string and a different extension hash on every visit is not the profile such systems were tuned to expect. The same defence that protects a user from cross-site correlation can render that user statistically unusual to a detection model, and unusual sessions are the ones that get scored down.

Scale determines how much of this actually bites. Cloudflare's Q3 2025 measurement placed Brave at 1.083% of global web traffic against Chrome's 66.282%. By user count the picture is larger: Brave passed 100 million monthly active users in October 2025 and reported 109 million by March 2026.

Much of that traffic was already dark. Brave blocks advertisements and third-party trackers by default through Shields, meaning those sessions largely do not appear in conventional analytics or ad requests in the first place. Research from Ad-Shield, based on more than 5 billion page views, put the population of users generating traffic invisible to measurement tools at close to 976 million in 2025, with a projection of 1.1 billion for 2026. The GPU change therefore hardens a cohort that was, for most advertising purposes, already out of reach. Its practical significance lies less in the immediate inventory affected than in what it signals about where default browser behaviour is heading.

Brave's own commercial position complicates any reading of the announcement. The company sells advertising against the same browsing sessions in which it blocks other parties' advertising, an arrangement that has drawn criticism since the ad network launched in January 2023. It also runs a search business, which gained a placement on Chrome's default engine list earlier this month, and a paid browser variant, Brave Origin, released on June 4, 2026 at 59.99 dollars, which strips 13 features while retaining Shields. Privacy engineering is, for Brave, both a public good and a product differentiator.

What the disclosure does not settle

Several questions remain open after the post.

No entropy figure accompanies the change, which makes it difficult to compare against Mozilla's published reduction in trackability. No breakage rate from the Nightly and Beta testing period is disclosed, so the compatibility claim rests on the company's assessment rather than published data. The crawl that produced the finding about fingerprinting-only use of the graphics interfaces is described as small, without further specification. And the WebGPU extension list, the analogue of the WebGL signal Brave chose to randomize, remains exposed pending future work.

There is also a structural limit the post acknowledges in passing. Graphics interfaces are one active research area among several, and each new capability added to the platform arrives with its own descriptors. Brave commits to extending coverage to new signals as they appear, which is a maintenance commitment rather than a resolution. The company closes by restating that it already blocks state-based tracking and known-dangerous scripts, describing the combined effect as "a win-win-win-win for privacy, security, performance and usability."

Timeline

Summary

Who: Brave Software, developer of the Chromium-based Brave browser. The engineering work is attributed to Rohit Agarwal, senior software engineer, and Peter Snyder, principal researcher, with the announcement written by Shivan Kaul Sahib, vice president of privacy and security. The changes affect tracking companies, anti-fraud vendors and measurement providers that read graphics hardware descriptors, alongside Brave's user base of more than 100 million monthly actives.

What: Three protections against graphics-driver fingerprinting shipped in version 1.93. The WebGL vendor and renderer strings are replaced with a single generic value identical across all Brave users, the WebGPU adapter descriptors are emptied, and the WebGL extension list is randomized per session, per site at the eTLD+1 level and per storage area. Randomization of WebGPU supported extensions is described as planned future work. No figure for the resulting reduction in identifiability was published.

When: The post was published on August 13, 2026, with a phased rollout described as taking place over the following several days. The protections had been running on the Nightly and Beta channels for several months beforehand.

Where: Desktop and Android builds of the browser, enabled by default. The announcement does not mention iOS. Verification is possible through the Electronic Frontier Foundation's Cover Your Tracks tool.

Why: Graphics interfaces disclose the exact vendor and model of a device's graphics processor, its driver, and its supported feature set, producing signals stable enough to identify a user without cookies or consent. Brave states that a crawl of top websites found most were calling these interfaces exclusively for fingerprinting. The change arrives as browser vendors and the largest advertising platform move in opposite directions on the technique, with Google having permitted advertiser use of device fingerprinting from February 16, 2025 while Mozilla, Apple and Brave have each tightened default restrictions.