Playwright screenshot returns noise: readback fix
page.screenshot() returned a full-frame PNG of colored static, tens of megabytes of it, because a canvas anti-fingerprint defense also rewrote the browser’s own screenshot readback. The noise was the same every time for a given seed and different for a different seed. The automation ran, the navigation succeeded, the file was written, and the file was garbage. The fix keyed the transform on a security boundary the engine already draws, exempting the privileged screenshot read while leaving every web-content read spoofed.
This is a closed bug in this product, written up because the shape of it is instructive.
The cause was a fingerprint defense that was doing its job a little too well, and could not tell the browser’s own screenshot machinery apart from a web page reading its own canvas. The fix was to teach it the difference. Here is the whole chain: the symptom, why a canvas defense reached a screenshot at all, and the principal split that fixed it without weakening the defense by a single byte.
The symptom: a screenshot that came back as static
The report was concrete. A script navigated, waited for the page to settle, and called page.screenshot(path="out.png"). The call did not error. The PNG was roughly 60 MB, which is already wrong for a normal page capture, and opening it showed uniform per-pixel noise across the entire frame with no page visible underneath.
Two details narrowed it immediately. First, the noise was stable for a fixed seed and changed with the seed, so it was not random corruption; it was deterministic and it was ours. Second, the page itself worked: text extraction, clicks and DOM queries all returned correct data. Only the pixels coming back through the screenshot path were wrong.
That combination points away from the network, the proxy and the page, and straight at a pixel transform that should only touch web content but was touching the capture as well.
Why a canvas defense reached the screenshot at all
To keep a canvas fingerprint from revealing the host OS, this build rewrites the pixels a web page reads back from a canvas. When a page draws to a canvas and then reads it, the bytes it gets are substituted with a per-seed value, so the resulting hash depends only on the seed and not on whether the underlying rasterizer is a Windows, Linux or macOS one. That is what keeps a canvas hash consistent with the rest of the spoofed machine. The mechanism is described in canvas fingerprint noise, and its cross-platform purpose in canvas and WebGL cross-platform consistency.
The trap is in the word “readback”. A screenshot is also a readback. When Playwright captures a page, the browser renders the current window into an offscreen surface and reads the pixels out, and internally that read went through the same code path as a page reading its own canvas. The pixel substitution did not distinguish the two, so it rewrote the screenshot pixels exactly as if the page had asked for them. The bigger the viewport, the bigger the noise PNG, which is where the tens of megabytes came from.
So the defense was not broken. It was correct and it was blind: it protected every readback, including the one readback that is the automation looking at its own page, which no detector ever sees and which therefore never needed protecting.
The fix: split the privileged reader from the web page
The distinction that solves this already exists inside the browser. A read that originates from web content runs under that page’s origin. A read that originates from the browser’s own machinery, such as the screenshot capture, runs under the browser’s trusted system context, the same privileged context that Firefox itself exempts from its built-in canvas fingerprinting protection. The two are not guesses about intent; they are a property the engine already carries on every readback.
The fix reuses exactly that boundary. Before substituting pixels, the canvas code now checks whether the read is coming from a privileged, trusted context (the browser’s own system context, or its internal chrome and resource origins) or from ordinary web content. Privileged reads are passed through untouched, so the screenshot is the real page, byte for byte. Web-content reads are substituted exactly as before, so the fingerprint stays spoofed. It is the same exemption boundary the browser’s own canvas protection uses, applied to our own transform rather than invented for it.
The whole fix is that one distinction:
| Read origin | Pixels returned | Who can trigger it |
|---|---|---|
| Privileged, trusted context (the screenshot capture, internal chrome and resource origins) | The real page, unchanged | Only the browser’s own machinery |
| Ordinary web content | Substituted per seed, so the hash matches the spoofed machine | Any page, including a detector |
The result, verified on the shipped binary at a fixed seed: screenshots are identical to a build from before the substitution existed, while a web page’s canvas read at that same seed returns the same spoofed bytes it did the release before. The screenshot got its page back and the fingerprint lost nothing.
Reproduce it: capture the page and check the bytes
The whole point of a seed is that a bug like this is reproducible rather than anecdotal. Fix the seed, capture, and assert two things at once: the screenshot is a sane size and it is a real image, not a wall of noise.
pip install invisible-playwright
from invisible_playwright import InvisiblePlaywright
with InvisiblePlaywright(seed=42) as browser:
page = browser.new_page()
page.goto("https://example.com", wait_until="networkidle")
png = page.screenshot() # bytes, no per-seed static in them
with open("capture.png", "wb") as f:
f.write(png)
# A full-frame noise PNG barely compresses and runs to tens of MB.
# A real page capture at this viewport is a few hundred KB.
print("screenshot bytes:", len(png))
assert len(png) < 5_000_000, "screenshot is suspiciously large - open it"
If you want to see both halves of the split in one run, read a canvas from inside the page as well. The screenshot stays clean while the in-page read stays spoofed and seed-stable, which is the behavior the fix guarantees:
from invisible_playwright import InvisiblePlaywright
def canvas_hash(page):
return page.evaluate("""() => {
const c = document.createElement('canvas');
c.width = 300; c.height = 150;
const ctx = c.getContext('2d');
ctx.textBaseline = 'top';
ctx.font = '16px sans-serif';
ctx.fillStyle = '#069';
ctx.fillText('readback split demo', 10, 10);
return c.toDataURL();
}""")
with InvisiblePlaywright(seed=42) as browser:
page = browser.new_page()
page.goto("https://example.com")
page.screenshot(path="clean.png") # the real page
first = canvas_hash(page) # spoofed, per-seed
with InvisiblePlaywright(seed=42) as browser:
page = browser.new_page()
page.goto("https://example.com")
second = canvas_hash(page)
assert first == second, "same seed must give the same canvas readback"
Open clean.png. Reading the log is not enough here; the original bug was caught by opening the PNG, because the byte count alone was merely surprising while the pixels were unambiguous. That habit is worth keeping for any capture, as the testing guide argues at more length.
What the fix does not change
Worth stating plainly, because a “we stopped rewriting pixels” headline invites the wrong conclusion. The anti-fingerprint substitution is unchanged for every read that a detector can actually make. A page that draws to a canvas and hashes the result gets the same per-seed value it always did, so a canvas hash still looks like the spoofed machine rather than the host that built the binary. Only the browser’s own privileged captures, which no page and no detector can issue, were exempted.
Nothing about the screenshot leaks the real machine either, because a screenshot is a picture of the page, not a probe of the host: it carries whatever the page rendered, and the surfaces a page could interrogate to unmask the host, such as the renderer string versus the actual render, are governed by their own defenses that this change does not touch. The split is narrow on purpose. It restores one trusted path and leaves every untrusted one exactly as spoofed as it was.
Conclusion
A screenshot that comes back as per-seed static is not flakiness and not a bad proxy. In this product it was a single, closed defect: a canvas fingerprint defense that protected every pixel readback, including the browser’s own screenshot capture, and rewrote the capture into noise. The fix was not to weaken the defense but to give it the distinction the engine already draws, between a privileged trusted read and a web-content read, and to exempt only the former. Screenshots are byte-for-byte clean, the fingerprint is byte-for-byte spoofed, and both were proven on the shipped binary at the same seed.
The general lesson outlives the bug: a defense that transforms output has to know who is asking, and the safest boundary to key on is one the platform already enforces rather than one you invent.
Short answers to the questions that lead here
Why does page.screenshot() return noise or a huge PNG? In older builds a canvas anti-fingerprint transform also rewrote the browser’s own screenshot readback, producing a full-frame per-seed noise image tens of megabytes in size. Update to a build with the privileged-readback exemption and the capture returns the real page.
Is the screenshot noise random or corruption? Neither. It was deterministic per seed, which is exactly what identified it as the fingerprint substitution rather than a decode or transport error.
Does fixing screenshots weaken the fingerprint? No. Only privileged, trusted reads (the screenshot path) were exempted. A web page’s canvas read is still substituted per seed, verified identical to the previous release at the same seed.
My clicks and text extraction work but the screenshot is wrong. Why? Because the page and the DOM were never affected; only the pixel readback path was. That split is the tell that points at a capture-side transform rather than the page.
How do I confirm the capture is clean? Fix a seed, capture, check the byte size is in the hundreds of KB rather than tens of MB, and open the PNG. A text log will not show you a wall of noise; the pixels will.
Can I still read a canvas for my own use in the page? Yes, and it returns a stable per-seed value. That read is web content, so it is spoofed by design; if you need faithful pixels for real work, capture them through a screenshot, which is the exempted path.
Sources
- This project’s patch notes for the canvas readback path, including the firefox-17 change that exempted privileged, trusted reads from pixel substitution, and the before/after verification on the released binary at a fixed seed.
- The engine’s own canvas fingerprinting protection, whose privileged-context exemption this fix reuses rather than reinvents.
See also: canvas fingerprint noise for what the substitution is protecting against, canvas and WebGL cross-platform consistency for why the transform exists at all, and the checklist for being detected on one site when the problem is the page rather than the capture.
Written while maintaining invisible_playwright, a Firefox patched at the C++ level driven by stock Playwright. This bug shipped, was caught by opening a PNG, and was closed by keying on a boundary the engine already draws.