Evidence. Reproduction. Repair.

Browser Busters Browser app testing & verified repairs

Failures reproduced. Repairs checked.

An AI-assisted Ten Thirty Two project testing our own browser-based tools: search and saved-list recovery, lost review notes, invisible CSV identifiers, startup races and slow rendering. Inspect the evidence, the repair and the limits. Historical browser security advisories are clearly attributed below.

Explore the case studies Open the evidence file
Original concept artwork: a hand uses a precision gauge to inspect layers of glass held in a browser-shaped workshop frame.
THE INSPECTION BENCHOriginal concept artwork. Findings are documented below.

New audit / Hand & Signal

Local doesn’t mean locked.

File defenses held in this targeted pass. A browser-saved copy was still readable from another path on the same site. See the evidence, the limits and practical ways to protect your work.

Read the workspace audit
The human and mechanical drawing hands from Hand & Signal.

Original artwork from the tested workspace. No new exploitable vulnerability was reproduced.

External service / Opinion

Outlook outage: what the evidence supports

Microsoft reports unavailable servers and pressure on remaining capacity. Our assessment separates the official report, recovery recommendations and unresolved cause. No verified Outlook fix or evidence of exploitation is claimed.

Read the Outlook recovery assessment ↗

Original findings / Explained

Find the failure you recognize.

For developers, testers and people who rely on browser-based tools. These case studies explain failures we reproduced in our own applications, how we repaired them and what the checks cannot establish.

  1. FMP-UNICODE-001

    How a split Unicode character cleared visible results

    A valid long source name could be cut through a Unicode character, breaking contact-link rendering. The repaired link keeps the listing and saved data usable.

    Repaired
  2. FMP-METADATA-001

    When nested source metadata interrupted a listing

    An object-shaped source identifier passed the listing boundary and broke optional attribution rendering. String checks now ignore unsupported metadata while keeping the listing usable.

    Repaired
  3. FMP-CATALOG-001

    When broken suggestions stop a typed search

    Malformed occupation suggestions disabled typed search. The repair validates the optional catalog before replacing it, so typing remains usable.

    Repaired
  4. FMP-RESTORE-001

    Recovering a saved list without overwriting it

    Malformed saved records interrupted startup and restoration. Recovery now preserves original storage, usable entries and the last valid cross-tab view.

    Repaired
  5. FMP-SOURCE-001

    When an inherited source name cleared visible results

    An unsupported source label matched an inherited object property. A supported-key check now keeps that metadata from interrupting result rendering.

    Repaired
  6. OCA-RECOVERY-001

    When a failed file load erases the previous review

    Follow the replacement state, compare the controls and see how the repair keeps existing work intact.

    Repaired
  7. SFC-IDENTITY-001 / MG-IDENTITY-002

    Invisible Unicode identifiers in CSV exports

    Two distinct validation gaps, exact identifier preservation and the limits of a character blacklist.

    Repaired
  8. MG-RESOURCE-001

    Why a small CSV produced 437,704 DOM elements

    Trace hidden source-detail rendering and compare the measured before-and-after boundary.

    Repaired
  9. FMP-RESPONSE-001

    When a malformed response erased good search results

    A successful request can still return unusable data. See the before-and-after checks that keep a completed search available.

    Repaired

02 / Findings

Separate the failure from its cause

These are our own test observations. The labels distinguish application issues, bounded negative tests, local browser restrictions and testing-tool effects.

FMP-UNICODE-001 / Application reliability · Repaired

How a split Unicode character cleared visible results

A valid long source name could be cut through a Unicode character, breaking contact-link rendering. The repaired link keeps the listing and saved data usable.

The demonstrated source path begins with a valid business name containing a character represented by a UTF-16 pair across the producer’s 180-code-unit cutoff. The fictional record passes the relevant source category and location filters and has no usable website, so the browser builds a Find contact link. Our own normalizer splits the character. We did not establish that a real provider currently supplies such a record or that any attacker controls it. The seeded saved-list case is a separate prerequisite, not proof that the baseline UI can normally save the already-failing record.

Repair: Build the fixed HTTPS contact-search URL with URL and URLSearchParams. An unpaired surrogate becomes a replacement character only in that outgoing query. Keep the original listing name, identifier and saved record unchanged. The same card renderer now works during search and saved-list restoration. This is a consumer repair: upstream truncation remains unchanged, and the page does not claim that all Unicode source handling has been corrected.

Read the reproduction, controls and limits · Try My People Release 7

FMP-METADATA-001 / Application reliability · Repaired

When nested source metadata interrupted a listing

An object-shaped source identifier passed the listing boundary and broke optional attribution rendering. String checks now ignore unsupported metadata while keeping the listing usable.

The candidate requires malformed nested location evidence for a supported source, with the surrounding attribution fields matching that source. No native producer was shown to create this object shape, and no production injection or attacker access was established. The same unsupported identifier with an unknown source key was ignored in the independent control; a canonical string identifier retained legitimate attribution.

Repair: Require primitive strings before validating location-evidence record IDs and dates. Apply the same check to secondary Ohio and Illinois location-source IDs and the Illinois checked date. Ignore unsupported optional metadata while retaining the underlying listing and its original saved record. Existing source, method, URL, identity and calendar checks still apply; valid source attribution remains visible.

Read the reproduction, controls and limits · Try My People Release 7

FMP-CATALOG-001 / Application reliability · Repaired

When broken suggestions stop a typed search

A fresh My People page received a malformed occupation catalog. The page said suggestions were unavailable and that a service could still be typed, but typing or submitting then failed before a search request was made. A valid catalog and a valid empty catalog each allowed one fictional search and one visible result.

Repair: Read the response within a 1 MiB byte limit, validate and detach the full catalog, and prepare its controls before committing it. Reject malformed or oversized suggestions while retaining the usable category state. A valid empty catalog remains a supported state, so typed search still works.

The unchanged-source run covered eight catalog shapes in each engine. Positive controls were a valid catalog and a valid empty catalog. Repaired cases require a completed typed search and no uncaught page error. Unit checks also exercise the actual 1,027-entry catalog, duplicate slugs, UTF-8 byte limits and cancellation of oversized streams.

Fictional local data reproduced the failure in Edge and Windows Playwright WebKit. No browser-engine flaw, attacker-controlled path or customer-loss measurement is established. Read the case and its limits · Try My People Release 6

FMP-RESTORE-001 / Application reliability · Repaired

Recovering a saved list without overwriting it

A synthetic saved list contained one valid person and a second record with a numeric phone or object-shaped listing type. Both records passed the old basic identity check. Rendering stopped after one card, while the saved count showed two; startup navigation did not initialize. A malformed cross-tab replacement and a pre-existing malformed history entry also interrupted restoration. Original local-storage bytes were not deleted.

Repair: Use a shared listing contract before rendering or restoring. At startup, show usable recovered entries while retaining the original storage bytes. Pause mutations when recovery is needed; explicitly clearing the entire saved list is a deliberate reset. Reject a malformed cross-tab replacement before changing the committed in-tab list. Reject malformed history snapshots before applying them. A storage event does not undo another tab’s write: the repair preserves this tab’s committed view, not an atomic cross-tab transaction.

Valid and empty stored lists remain usable. Repaired checks cover mixed and all-invalid records, preservation of raw storage, blocked implicit overwrites, explicit clearing, cross-tab rejection, malformed history and ordinary filters with Back, Forward and refresh. These are local-state recovery checks, not account authorization tests.

Fictional local data reproduced the failure in Edge and Windows Playwright WebKit. No browser-engine flaw, attacker-controlled path or customer-loss measurement is established. Read the case and its limits · Try My People Release 6

FMP-SOURCE-001 / Application reliability · Repaired

When an inherited source name cleared visible results

My People showed one fictional result and saved it. A replacement home-health record used an inherited object-property name as its location source, with the listing state and the location-evidence method and URL absent. The plain object lookup returned an inherited value, and rendering later tried to use a missing validation rule. The replacement left zero visible cards. The saved shortlist remained intact.

Repair: Require the source label to be an own property of the application’s fixed source registry before consulting its validation rules. Unsupported metadata is ignored. The listing can still render, and supported source attribution is preserved.

Three inherited names and one ordinary unknown-source control were checked in each engine on exact baseline source. The six inherited-name cases changed one visible result to zero; both unknown-source controls retained one. Repaired checks require one visible result and preserved saved state for every case.

Fictional local data reproduced the failure in Edge and Windows Playwright WebKit. No browser-engine flaw, attacker-controlled path or customer-loss measurement is established. Read the case and its limits · Try My People Release 6

FMP-RESPONSE-001 / Application reliability · Low severity · Repaired

A malformed search response erased the completed results.

My People displayed 12 fictional listings. A later successful HTTP response containing a null listing name left zero visible results and disabled the app’s history navigation. The saved shortlist stayed intact. Both Microsoft Edge 157.0.4341.46 and Playwright WebKit 26.5 on Windows reproduced the failure.

Repair: validate the candidate before committing it. The same two browser checks now retain all 12 results and show a retry message. Six targeted regressions cover malformed response shapes and valid empty results.

An ordinary failed request retained the results in both versions. Tested markup-shaped names stayed text and unsafe website links were blocked. These fictional local checks establish an application defect, not a browser-engine vulnerability, a customer incident or an attacker’s ability to provide the response. Physical mobile devices remain untested.

Read the search-response case study · Try My People Release 5

OCA-RECOVERY-001 / Application reliability · Low severity · Repaired

A failed sample load removed an existing review.

In Order Confirmation Auditor, trying another fictional sample cleared the current file pair, comparison and notes before the sample was ready. A controlled request failure left that in-tab work unavailable. Original files and reports already downloaded were not deleted.

Measured result: the failure reproduced in Edge and Playwright WebKit, with another independent Edge reproduction. Changing only the selector preserved the existing report; successfully loading a replacement correctly started a new review.

Repair: the current workspace stays intact while a candidate sample is checked. Comparison, review editing and exports wait. Failure retains the prior review; success replaces the pair together. Later source, tolerance or clear actions cancel pending work and prevent old results from returning.

All 23 targeted local regression cases passed after the repair, including malformed samples and canceled or superseded requests. Open Order Confirmation Auditor Release 3.

Understand the recovery test and its limits

On an authorized local copy with fictional files, annotate a comparison and download its JSON. Change the sample selector without loading: the report should remain unchanged. Then fail a replacement sample request locally and check that the original pair hashes and annotations are still available after failure. A successful replacement should begin unreviewed.

The environments were Microsoft Edge 157.0.4341.46 and Playwright WebKit 26.5 on Windows. The matrix includes controlled old-handler dispatches as robustness checks, not an attacker capability. This is an application recovery defect, not a browser-engine vulnerability. We did not measure real customer loss or how often this occurred. Firefox and physical Safari/mobile devices remain untested. See OCA-RECOVERY-001 and recoveryRegressionSummary in the evidence file.

Read the review-recovery case study

APP-STARTUP-001 / Application reliability · Low severity · Repaired

A visible file control was not always a ready file control.

Holding back the application module exposed an interaction gap in four of our CSV tools. Three sample buttons ignored early clicks. Four file inputs held a selected synthetic file, but application status did not reflect it.

Measured result: 21 lost actions in 24 controlled checks across Edge, Chrome and Playwright WebKit. Margin Guard’s already-protected demo button supplied the three passing negative controls. Repeating every action after initialization worked.

Repair: intake and policy controls now wait for completed initialization. In the matching 24 repaired checks, every tested control stayed unavailable before readiness and worked afterward. Existing comparison and export prerequisites remain in place.

Verified repairs: Margin Guard Release 6, Supplier File Converter Release 3, Supplier Quote Comparator Release 2 and Order Confirmation Auditor Release 2.

Reproduce the boundary, understand the limit

On an authorized local copy with synthetic files, hold the application module request after the HTML appears. Try a sample or select a file; then release the request and inspect the application’s loaded-file status. Repeat after initialization as a control. Our file selections used browser automation, not a physical operating-system picker.

The startup checks ran in Microsoft Edge 157.0.4341.46, Google Chrome 156.0.8078.48 and Playwright WebKit 26.5 on Windows. The controlled delay establishes that this failure was possible; it does not measure how often a visitor encountered it. This is an application readiness defect, not evidence of a browser-engine vulnerability, data disclosure or account compromise. See APP-STARTUP-001 and startupRegressionSummary in the evidence file.

Application data integrity · low severity · fixed

Invisible product IDs reached reviewed exports.

A Margin Guard SKU containing only U+200B ZERO WIDTH SPACE could produce a price-change recommendation and appear in the downloaded changes CSV after manual approval. Supplier File Converter blocked that character, but a SKU containing only U+3164 HANGUL FILLER passed its validation and approved import export.

Both paths were reproduced with fictional local files in fresh Edge contexts. The visitor had to choose the files and explicitly approve the export. We observed invalid identity data in the exported CSVs; no live inventory change, financial loss, data disclosure or code execution was observed.

Repair: check for meaningful identity content as well as unsafe formatting controls, while keeping the original source evidence. Margin Guard Release 3 and Supplier File Converter Release 2 block the tested invisible-only and directional-control cases. Six legitimate leading-zero, Arabic, CJK and joiner-containing identities survive the paired conversion, changes and rollback exports unchanged.

The new regressions failed on the original code. After repair, all 28 Margin core checks, 12 Converter core checks and three targeted browser cases passed. This is a conditional application defect; it does not establish a browser-engine vulnerability or guarantee how every Unicode character renders in every font.

Reproduction boundary and repair evidence

Use synthetic data only. The pre-repair Margin case required the same invisible SKU in both cost and price files. The Converter case required an explicit column mapping and review. The repaired controls keep these rows available as evidence but prevent approved output. A finite blacklist alone missed the Hangul filler case; deleting characters would silently change identity, so the repair preserves the raw value and requires correction.

Finding records MG-IDENTITY-002 and SFC-IDENTITY-001 are in the sanitized evidence file. Original fixtures and runnable reproducers remain in the private test workspace. Publication contains observations and repair details, with no executable exploit payload.

Read the invisible-identifier case study

Application availability · low severity · fixed

A small CSV could create a huge hidden page.

Margin Guard accepted a synthetic 500-row cost file with 410 columns. Although it was only 226,863 bytes, rendering created 437,704 page elements, including 425,524 inside closed details. One measured main-thread task lasted 1,099 ms.

This required selecting a crafted file in the app. We observed temporary interaction blocking, with no crash or data disclosure in that test. It is an application resource-amplification issue, not a newly discovered browser-engine flaw.

Repair: limit each source to 100 columns and create detailed source evidence only when a row is opened. The previous wide file now rejects. The comparable accepted 500-row, 100-column case fell from 127,704 to 12,705 initial elements, with no evidence bodies until requested. Timings and element counts describe this controlled environment, not a universal performance guarantee.

Read the CSV rendering case study

Adversarial application checks · passed

Test the controls that protect the workflow.

In the cases run, markup-like input stayed text, prototype-shaped keys did not alter object prototypes, stale file reads could not restore cleared results, and changing inputs revoked approvals and prepared downloads. Formula-risk identifiers were blocked from approved imports.

Both full applications rejected framing. The calculator’s intentional same-origin embed worked while a different-origin frame was denied. We observed no external data requests or persistent browser storage in these synthetic workflows.

These negative results cover the tested inputs and policies. They do not prove that every injection, race, extension or network configuration is safe.

Environment restriction

WebKit rejected a local test port.

The WebKit test environment refused local port 4190 before an application request reached the server. The same application bytes worked on port 4174.

The observed failure was local navigation, not evidence of an application defect. WebKit’s official portAllowed source lists 4190 among blocked ports. Inspect the source and search for 4190.

Test instrumentation

Screenshot capture produced a policy warning.

Our screenshot instrumentation injected styling that WebKit reported as a Content Security Policy (CSP) violation. A control run without screenshot capture completed all eight workflow checks without those errors.

The strict application policy was retained. Captured screenshots were reviewed separately; the warning is recorded as a testing-tool effect, not a discovered browser vulnerability.

Application issue · fixed

A small companion link needed more touch space.

At a 320px viewport, a companion navigation link measured 42.09px high. Its target was increased to at least 44px and checked again.

This was a layout defect we could correct directly. The measured target size is evidence for that control, not a claim of complete accessibility certification.

Data integrity · verified

The file checksum includes the original BOM.

The live supplier-to-margin round trip preserved SKU 00123. A synthetic UTF-8 file with a byte order mark (BOM) matched its displayed SHA-256 checksum to the exact uploaded bytes, including the BOM.

This verifies file identity in that test. It does not establish that the contents of a supplier’s real file are correct.

Source for the observations and test scope: our sanitized test evidence (JSON).

01 / Coverage

Where the checks ran

The original two-tool checks below cover Margin Guard and Supplier File Converter. The separate four-tool startup experiment is described in APP-STARTUP-001; the auditor recovery checks used Edge and WebKit and are described in OCA-RECOVERY-001. Windows test environments with fresh automation profiles passed the listed checks at 320, 390, 768 and 1440 CSS pixels. Results cover these versions and scenarios; viewport emulation is not a physical phone test.

Browser environments and the limits of their test coverage
EnvironmentResultScope
Microsoft Edge 157.0.4341.46CheckedBoth apps: synthetic-file workflows and four viewport widths.
Google Chrome 156.0.8078.48CheckedBoth apps: synthetic-file workflows and four viewport widths.
Playwright WebKit on Windows 26.5CheckedBoth apps: workflows and four viewport widths. This is not Apple Safari.
FirefoxNot yet testedTest runtime could not be installed in this environment; no Firefox result is claimed.
Physical Safari, iPhone and AndroidNot yet testedDevice behavior, native file pickers and mobile download handling remain unverified.
Brave, Opera, other browsers and embedded webviewsNot yet testedNo result is inferred from a shared engine. Extensions and personal browser-session configurations were not tested.

Playwright explains that its WebKit build is patched for testing and differs from branded Safari. Read the official browser guidance.

03 / Repeat the workflow

Use the fictional sample

No real catalog is needed. The expected counts below describe bundled synthetic data, not customer records or business results.

  1. Start in Supplier File Converter. Choose Try a sample file and explicitly set the review date to 2026-11-10. Inspect the loaded example mappings, then validate. Expect 11 rows: 4 ready and 7 needing review.
  2. Inspect the blocked rows. Look for duplicates, missing freight, an ambiguous unit and the unsafe identifier. Confirm that original source lines remain visible. Download the row-review CSV before approving any import.
  3. Approve the limited export. Review the ready rows and separately approve exporting only those four. Check that SKU 00123 retains its leading zeros. In a spreadsheet, import the SKU column as Text.
  4. Check the companion result. Open Margin Guard with that exported cost file and the converter’s companion prices CSV. Use audit date 2026-11-10, maximum cost age 90 days and target margin 25%. The four records should pass blocking gates; 00123 should show 20.00% gross margin.
  5. Check that old approvals expire. Change a mapping or review date, or clear the converter. Its earlier result and prepared download must disappear. Neither application should update a live inventory system.
Repeat the checksum and layout checks

Use a synthetic UTF-8 CSV with an initial BOM. Compute SHA-256 from its original bytes using a local file-hashing tool, load it in the converter, and compare the full value with the conversion record. A BOM-bearing file and a file without that BOM have different bytes and should not be assumed to share a checksum.

At each listed viewport width, inspect the initial page, expanded source evidence and approved-download state. Check keyboard focus, horizontal overflow and touch targets. Record the browser version and distinguish browser navigation, application behavior and test-tool errors. Repeat any screenshot-only warning without capture before attributing it to the app.

Bounded controls / No engine finding

Check the isolation boundary, too.

Six paired local scenarios per engine passed in Microsoft Edge 157.0.4341.46 and Windows Playwright WebKit 26.5: 12 of 12 scenarios. They compared same-origin DOM access with another loopback port, same-origin storage with another origin and a separate browser context, and restrictive with permissive sandbox fixtures. The actual My People link supplied the opener-isolation check, and the application CSP was tested locally.

Frame-denial configuration was mirrored on a local server; this is not a live-header verification. Generic platform fixtures and application controls are distinct evidence. No browser-engine flaw was demonstrated, and these bounded controls do not establish universal safety or account isolation. The original WebKit test-harness timing error was retained and corrected before the completed run; it was not an isolation bypass.

Inspect the platform control summary and application evidence

Field intelligence / Vendor disclosures

Real flaws. Exact boundaries.

These selected cases span Chromium, Gecko and WebKit. They are published vendor findings, not new discoveries by Browser Busters or a complete vulnerability catalog. Advisory dates describe the source events. A listed fix is a historical security threshold, not a recommendation to install an old browser.

Match the product, platform and release channel. A patch announcement does not prove that a device installed it or restarted. Passing our application tests does not establish browser patch status. Use the latest supported update offered for your product.

Chrome / XML · Vendor severity: high

CVE-2026-125350

Google’s November 9, 2026 release reports exploitation of this use-after-free flaw. Stable fixes are 157.0.8103.50/.51 on Windows and Mac and 157.0.8103.50 on Linux. The separate Extended Stable fix is 156.0.8078.96 on Windows and Mac.

The CVE description refers to Chrome before 157.0.8103.50 and code execution inside the sandbox. The explicit Extended Stable backport means that one global version cutoff would misclassify a fixed branch. This record alone does not demonstrate a sandbox escape.

Google Stable advisory · Extended Stable backport · CVE record

Chromium / V8 · Vendor severity: medium

CVE-2026-87491

Google’s September 8, 2026 note reports an exploited out-of-bounds write. Its Chrome fix is 153.0.8010.36/.37 on Windows and Mac, or 153.0.8010.36 on Linux. The CVE description identifies Chrome before 153.0.8010.36.

Microsoft’s separate September 4 entry names an Edge fix at 152.0.4191.66. These product build numbers are not interchangeable. Microsoft’s November 9 notice mentions recent exploits without naming a CVE; we cannot assign that notice to a particular flaw.

Google release note · Microsoft security release notes · CVE record

Chrome / Skia · Source discrepancy

CVE-2026-3909

Google corrected its March 12, 2026 notes to remove this flaw from that update’s fixes. The March 13 note explicitly supplies the fix in 146.0.7680.80 on Windows, Mac and Linux and reports exploitation.

The retrieved CVE record still describes versions before 146.0.7680.75. We preserve that discrepancy: the earlier cutoff does not clear .75–.79. The corrected vendor release supports the fix statement; the complete affected range remains unresolved here.

Corrected earlier note · Verified fix release · CVE version record

Firefox / WebAssembly · Vendor severity: critical

CVE-2026-15718

Mozilla identifies an invalid-pointer flaw, fixed in Firefox 152.0.6 in its July 14, 2026 advisory and Firefox ESR 140.13 in its July 21 advisory.

Those notices acknowledge public exploit code while reporting no known attacks at the time. Public code and observed attacks are different evidence. The notices do not establish today’s attack activity, a complete affected-version history or the patch state of an installed browser.

Firefox advisory · Firefox ESR advisory

Safari / WebKit · Linked disclosure

CVE-2025-43529 & CVE-2025-14174

Apple’s Safari 26.2 release for macOS Sonoma and Sequoia, released December 12, 2025, addresses linked use-after-free/code-execution and memory-corruption issues. The Safari CVE-2025-43529 record identifies versions before 26.2.

Apple describes possible sophisticated, targeted exploitation on iOS versions before iOS 26. That statement does not establish attacks on every Mac. Safari, iOS and other WebKit products require their own platform advisories; our Windows WebKit test build cannot clear them.

Apple Safari security advisory · Safari CVE record

What these records cannot tell us

Vendor reports establish what was disclosed. They do not measure how often attacks succeeded, prove that a particular person was targeted, or certify every product using the same engine. CISA catalog inclusion supports a known-exploitation classification, not a finding of local compromise. Absence from that catalog does not prove a flaw is unexploited.

We did not run browser exploits, scan third-party systems, inspect personal browser profiles or test extensions. The application security checks in this report used our own code and synthetic files. There is no universal “all browsers safe” result.

The crew / Four responsibilities

One team. A check on every claim.

These are roles in our AI-assisted workflow. Each finding needs a source, a repeatable check and a stated limit before it earns a place here.

  1. Scout

    Find the original advisory. Check dates, versions and what the author actually reported.

  2. Breaker

    Challenge our authorized applications with bounded, synthetic cases. Keep real people’s data out.

  3. Verifier

    Repeat the result, run a control and separate application failures from test-tool effects.

  4. Builder

    Repair confirmed defects. Add a regression check and verify the released result.