Two issues fixed:
1. autoDetectPII only skipped identity values (names, emails, etc.) but not
mapping values. After clicking "+" to map e.g. an IP address, the fake
substitute IP was immediately re-detected as PII, causing the notification
to reappear. Now autoDetectPII accepts opts.mappings and adds both real
and substitute values to the skip set.
2. The local mappings array was only updated after the async background
round-trip completed. The 150ms re-scan could fire before that, missing
the new mapping. Now an optimistic temp mapping is added to the local
array immediately so the re-scan already knows to skip both values.
https://claude.ai/code/session_01RfzvB5sHah326acr8Xa7Jn
The plus button handler in the pre-send PII warning had two bugs:
1. It read/wrote mappings directly via the storage bridge (api.storage.local),
bypassing the Storage module's encryption layer. When at-rest encryption was
enabled, getStorageData returned an encrypted blob instead of an array,
causing .push() to throw a TypeError that silently aborted the handler —
the mapping was never saved and replaceInInput never ran.
2. Unlike the ignore button which immediately removes the DOM item, the plus
button relied on a re-scan at 150ms to dismiss the notification. If
replaceInInput didn't stick (e.g. React-controlled inputs), the re-scan
found PII again and the notification persisted.
Fix: Route mapping creation through the background service worker via a new
'add:mapping' message handler (which uses Storage.addMapping with proper
encryption support), and immediately dismiss the notification item from the
DOM like the ignore button does.
https://claude.ai/code/session_01RfzvB5sHah326acr8Xa7Jn
If anything in Silent Send's substitution logic throws (bad mapping,
regex error, unexpected body format, etc.), the exception was propagating
up to the caller instead of falling through to the original fetch. This
broke Claude Code's web interface (claude.ai/code) when a bad mapping
caused a substitution error mid-request.
The inner try/catch on JSON.parse only covered the JSON path — the outer
logic (hasSubstitutions, processBody, notifySubstitutions, etc.) had no
protection. Added a top-level try/catch that catches any unhandled error
and falls through to originalFetch, logging a warning so the error is
still visible in the console.
https://claude.ai/code/session_01CwcZK8nqL8pyBH9AxDs9qo
A mapping imported with real: " " (single space) passed the !m.real
guard (space is truthy), producing new RegExp(" ", "gi") which matched
every space character in the DOM and highlighted them all.
- Added .trim() to all real/substitute guard conditions in content.js
(inline substitute, reveal functions) and substitution-engine.js so
whitespace-only values are treated as empty and skipped
- Added the same guard in highlightMatches so the CSS Highlight API
never creates ranges for blank search terms
- Added trim + blank-check to the bulk import path in options.js so
whitespace-only real values are dropped at import time rather than
saved to storage
https://claude.ai/code/session_01CwcZK8nqL8pyBH9AxDs9qo
The × close button only hid the warning for the current send — the same
detection reappeared on every subsequent send with no way to silence it.
Added an Ignore button to each item in both the on-send warning
(showAutoDetectWarning) and the pre-send input scanner warning
(showPreSendWarning). Clicking Ignore:
- Adds the specific value to an in-memory Set (ignoredDetections)
- Persists it to ss_ignored_detections in storage
- Removes that item from the current warning; closes the popup if empty
On page load, ss_ignored_detections is read from storage so ignores
survive page reloads. Ignored values are also filtered before
auto-redaction in the fetch interceptor, so they are not redacted either.
https://claude.ai/code/session_01CwcZK8nqL8pyBH9AxDs9qo
Two fixes:
1. grok.com missing from manifest
Grok moved from grok.x.ai to grok.com. The extension wasn't activating
on grok.com at all (no content script injected). Added to host_permissions
and content_scripts in both manifests.
2. Fetch hook bypassed on React/Next.js sites
content.js was injected async (after awaiting the document-scanner),
so by the time it replaced window.fetch, React/Next.js had already stored
a reference to the original. All fetch calls from framework code bypassed
the hook entirely.
Fix: load early-hook.js as a "world": "MAIN" content script at
document_start. This runs synchronously before any page JavaScript,
captures the real fetch in window.__ssOriginalFetch, and replaces
window.fetch with a lightweight wrapper. When content.js eventually
loads, it sets window.__ssInterceptFetch (the real substitution logic)
and window.__ssReady = true, activating the wrapper. React's stored
fetch reference now routes through the interceptor.
Falls back to direct window.fetch replacement if early-hook.js somehow
didn't run (e.g. older browser without world: MAIN support).
https://claude.ai/code/session_01CwcZK8nqL8pyBH9AxDs9qo
The fetch/XHR interception correctly checked settings.enabled, but
several other features ignored it — highlights, reveal mode, and
PII auto-detect all continued running after disabling. This adds
settings.enabled checks to the MutationObserver, highlightMatches,
auto-detect input/paste listeners, reveal mode, and cleans up
visual artifacts (highlights, warnings) when the extension is
toggled off.
https://claude.ai/code/session_015TEttQgcq5FALLKLb3uEW8
- document-scanner.js: Add missing await on _inflateSync() call — FlateDecode
streams were not being decompressed because the async function was called
without await, causing decompressed to hold a Promise instead of data
- content.js, storage.js, options.html, options.js: Gate proper noun / capital
letter detection behind a new detectProperNouns setting (default off); was
previously always-on causing noisy false-positive warnings on AI thinking
output and common capitalized words
- Bump version 0.9.23 → 0.9.25
https://claude.ai/code/session_01ReZUeR1nrYzJeX7fTqwXMw
The old unrevealInElement relied on a WeakMap (originalTexts) to restore
text nodes to their pre-reveal state. This broke when SPA frameworks
(React) re-rendered the DOM while reveal was active — new text nodes
containing real values had no WeakMap entry to restore from, so real
data stayed visible after turning reveal off.
Fix: unrevealInElement now actively reverse-replaces real values back
to their substitute counterparts using the reveal pairs, matching the
same approach revealText uses in the forward direction. This works
regardless of DOM re-renders or streaming content changes.
https://claude.ai/code/session_01NNBEPuXMFGWezJb1f958nL
- Inject document-scanner.js into page world via injector.js (as module,
sets globalThis.DocumentScanner)
- Add FormData interception to fetch hook: scans File/Blob entries through
DocumentScanner.processUpload(), also substitutes string fields
- Add document-scanner.js to web_accessible_resources in both manifests
- Bump version to 0.9.23 in package.json, manifest.json, manifest.firefox.json
https://claude.ai/code/session_01NNBEPuXMFGWezJb1f958nL
Going back to a known-good baseline. This version had:
- Working reveal mode with CSS Highlight API
- Working substitution (fetch + XHR hooks)
- Smart patterns (names, emails, phones, usernames)
- Encryption/sync (password, TOTP, WebAuthn)
- Multiple identity profiles
- Activity log
- Secret scanner
- Auto-detect PII warnings
- Pre-send PII detection
Kept current manifests (UUID, data_collection_permissions, version).
No renames applied — uses original naming (secretScanning, PPI, etc).
Will re-apply renames and new features from this working base.
https://claude.ai/code/session_01KF4i7Ra7zCEDskxDBaNtcT
The data-ss-config attribute on the script tag was being removed
(via script.onload) before content.js could read it. Changed approach:
inject config as a separate <script type="application/json" id="ss-config-data">
element that persists in the DOM until content.js reads and removes it.
This eliminates the race condition between script execution and onload
removal. Bump 0.9.20.
https://claude.ai/code/session_01KF4i7Ra7zCEDskxDBaNtcT
Reverted content.js to commit 9a11896 (the last version where reveal
mode worked) and re-applied only the renames:
- PPI → PII
- secretScanning → autoRedact
- SECRET_PATTERNS → REDACT_PATTERNS
- scanAndRedactSecrets → runAutoRedact
- category: 'secret' → category: 'redact'
- Added customRedactPatterns support to runAutoRedact
This restores:
- Individual word storage in sessionSubstitutions (needed for reveal)
- fetch(Request) handling (not just fetch(url, options))
- Blob/ArrayBuffer/URLSearchParams body conversion
- isInNonChatArea scoping (narrowed version)
- Proper noun detection gated by settings.detectProperNouns
- Mapping values added to configured skip set
https://claude.ai/code/session_01KF4i7Ra7zCEDskxDBaNtcT
The class-based selectors ([class*="header"], [class*="Header"],
[class*="nav-"], etc.) were matching Claude.ai's chat content area
elements, preventing reveal from running on any response text.
Narrowed to only structural elements: nav, aside, [role="navigation"],
[role="complementary"], [data-sidebar]. Removed HEADER/FOOTER from
SKIP_REVEAL_TAGS since sites use these tags inside chat layouts.
https://claude.ai/code/session_01KF4i7Ra7zCEDskxDBaNtcT
The smart engine records individual name parts ("Ademo"→"John",
"Demo"→"Smith") AND the combined form ("Ademo Demo"→"John Smith")
in sessionSubstitutions. The catch-all in buildRevealPairs was
adding all of them, causing partial replacements that corrupted
the DOM and made the cache oscillate between 4 and 0 pairs.
Fix: skip session entries whose key is a substring of a longer
entry (e.g. "ademo" is part of "ademo demo"). Only the combined
form gets added as a reveal pair.
Also: remove debug logging, improve cache with size tracking.
https://claude.ai/code/session_01KF4i7Ra7zCEDskxDBaNtcT
Root cause found: commit 9a11896 (v2.0.14) added isInNonChatArea()
and expanded SKIP_REVEAL_TAGS with NAV/ASIDE/HEADER/FOOTER, but that
commit was on a branch that was never merged into main. When we started
our work from main, these fixes were missing.
Restored:
- isInNonChatArea() function — prevents reveal from touching sidebars,
navigation, headers, footers, and other non-chat UI
- SKIP_REVEAL_TAGS expanded with NAV, ASIDE, HEADER, FOOTER
- isInNonChatArea checks added to revealInElement, unrevealInElement,
and highlightMatches (both element-level and walker-level)
https://claude.ai/code/session_01KF4i7Ra7zCEDskxDBaNtcT
Two bugs found:
1. Individual name parts ("Ademo"→"John", "Demo"→"Smith") were added
as reveal pairs, causing partial replacements that corrupted the DOM.
The smart engine sends combined forms ("Ademo Demo"→"John Smith")
which the catch-all already handles. Removed individual name entries
from buildRevealPairs — only emails, usernames, hostnames, phones
are matched individually.
2. Cache invalidated on every ss:config-updated (including settings-only
changes like reveal toggle). Now only invalidates when mappings or
identity actually change.
https://claude.ai/code/session_01KF4i7Ra7zCEDskxDBaNtcT
If reveal mode was saved as ON in settings, the page loaded with
revealMode=true and prevRevealMode=true. The checkRevealToggle
function only fires on transitions (off→on or on→off), so the
reveal interval was never started. Added initialization check
that starts the reveal interval immediately if revealMode is
already true at page load.
https://claude.ai/code/session_01KF4i7Ra7zCEDskxDBaNtcT
PII (Personally Identifiable Information) is the correct standard term.
Renamed all instances — comments, UI labels, variable names, function
names (autoDetectPPI → autoDetectPII, scanInputForPPI → scanInputForPII,
ppiWarnings → piiWarnings), CSS comments, README, and options page.
Added Ko-fi donation section to README and Options footer. Tone: no
obligation, no warranty, no influence on updates — just appreciation.
https://claude.ai/code/session_01KF4i7Ra7zCEDskxDBaNtcT
Users can now define custom regex patterns for proprietary token formats,
internal URLs with keys, or any secret the built-in scanner doesn't cover.
Patterns are added/toggled/removed from the Options page and apply to both
the live interception (content.js) and the Test tab (popup.js).
Renamed all user-facing "Secret scanning" labels to "Auto Redact" across
popup and options. Internal variable names (secretScanning, SecretScanner)
kept for backwards compatibility with stored settings.
https://claude.ai/code/session_01KF4i7Ra7zCEDskxDBaNtcT
Fixed safeHTML in all three files — template.content.childNodes and
doc.body.childNodes are live NodeLists that shrink as nodes are moved.
Using Array.from() to create a static copy before spreading into
replaceChildren().
Expanded proper noun common words filter with ~500 additional verbs,
nouns, and adjectives (generate, design, manage, process, account,
button, dashboard, etc.) to prevent false PPI flags on titles,
headings, and UI button labels.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Massively expanded the common words filter — added ~500 common verbs,
nouns, and adjectives that frequently appear capitalized in titles,
headings, UI buttons, and instructions. 'Generate Design', 'Create
Account', 'Search Filter', etc. are no longer flagged as PPI.
Updated both content.js and auto-detect.js with the same word list.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
sessionSubstitutions stored keys as toLowerCase() for lookup, but the
lowercase key was also used as the substitute value when building
reveal pairs. This caused unrevealText to replace "John Smith" with
"ademo demo" instead of "Ademo Demo".
Fixed by storing both the original-case substitute and the original
real value in sessionSubstitutions. The lowercase key is still used
for lookup, but the preserved-case value is used for reveal pairs.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
When reveal mode was turned OFF, text stayed showing real data
(e.g. "John Smith") instead of reverting to the AI's actual text
("Ademo Demo"). The unreveal function relied on a WeakMap of
original text nodes which lost entries when DOM nodes were replaced
during streaming responses.
Fixed by replacing WeakMap-based restore with active reverse
replacement: unrevealText() replaces real→substitute (the opposite
direction of revealText). This is reliable regardless of DOM changes.
Removed originalTexts WeakMap entirely.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
1. Removed data_collection_permissions from Firefox manifest — it
requires Firefox 140+ but we target 128+. Only produces a warning
when missing, not a blocking error.
2. Fixed JS syntax error in content.js line 450 — two extra closing
braces from a bad merge in the proper noun detection function
caused a parse error that blocked Mozilla validation.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Replaced all 31 innerHTML assignments across popup.js (12),
options.js (17), and content.js (3) to pass Mozilla AMO linter.
Each file gets a safeHTML(el, html) helper:
- popup.js/options.js: DOMParser-based (extension page context)
- content.js: <template> element pattern (page world context)
Empty innerHTML clears replaced with el.replaceChildren().
innerHTML += replaced with createElement + safeHTML + appendChild.
All event listener bindings after HTML rebuilds remain functional
since safeHTML uses replaceChildren() which populates DOM
synchronously before querySelectorAll + addEventListener calls.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
The X icon looked like a dismiss/close button (temporary). Changed to
a small 'IGNORE' text link that clearly communicates the action is
permanent — the value will never be flagged again.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Proper noun heuristic:
- Changed from matching any single capitalized word to requiring
TWO OR MORE consecutive capitalized words (e.g. "Acme Corp")
- Single capitalized words at sentence starts were causing massive
false positives — every sentence starts with a capital letter
- Minimum 5 characters total and 2 proper words required
Ignore button:
- Each PPI warning item now has an X (ignore) button alongside
the + (add mapping) button
- Ignored values are persisted to storage (ss_ignored_ppi) so
they stay dismissed across page reloads
- Ignored values are skipped in both pattern detection and
proper noun detection
- Clicking ignore removes the item from the warning and re-scans
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Reveal mode's text replacement was walking ALL text nodes on the page
including contenteditable elements (the chat input box on Claude.ai,
ChatGPT, etc.), overwriting the user's typed text with garbled
revealed content.
Fixed by adding contenteditable checks to:
- revealInElement() — returns early if element is contenteditable
- unrevealInElement() — same
- highlightMatches() — TreeWalker filter rejects contenteditable nodes
- MutationObserver — skips addedNodes and characterData mutations
inside contenteditable elements
All four code paths now use parent.closest('[contenteditable="true"]')
to detect and skip chat input areas.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Removed ZIP read/write complexity. All document formats now extract
text and upload as plaintext — the AI extracts text from files anyway,
no need to preserve formatting in a file the user never gets back.
Added support for: DOC, XLS (old binary), ODT, ODS, ODP (OpenDocument),
PPTX (PowerPoint), RTF.
Extraction methods:
- PDF: content stream operators (Tj, TJ)
- DOCX/XLSX/PPTX/ODT/ODS/ODP: ZIP XML text extraction with DEFLATE
- DOC/XLS: UTF-16LE and ASCII text run extraction from binary
- RTF: strip formatting commands, keep text
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
DOCX and XLSX files now get actual text substitution, not just
scan-and-warn:
- Reads the ZIP structure (DOCX/XLSX are ZIP files of XML)
- Decompresses DEFLATE entries via DecompressionStream API
- Finds XML text content between tags
- Applies substituteAll() to each text node
- Rewrites modified entries as uncompressed (stored) in new ZIP
- Unmodified entries preserved with original compression
- Rebuilds central directory and end-of-central-directory record
DOCX targets: word/document.xml, word/header*.xml, word/footer*.xml,
word/comments.xml, word/endnotes.xml, word/footnotes.xml
XLSX targets: xl/sharedStrings.xml, xl/worksheets/sheet*.xml
Formatting, styles, images, formulas all preserved — only text
content between XML tags is modified. Preview mode still shows
findings before upload for user confirmation.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
New module: document-scanner.js
- PDF: extracts text from content streams, scans for PPI, converts to
sanitized plaintext for upload (PDFs can't be reliably edited in-place)
- DOCX/XLSX: extracts text from XML entries, scans for PPI
- Text files: direct string substitution
- Handles scanned/image-only PDFs gracefully (skips with message)
Fetch interceptor (content.js):
- Now intercepts FormData uploads (file uploads) in addition to JSON
- Processes each file through document scanner
- Preview mode for PDF/DOCX/XLSX: shows PPI findings with counts,
user clicks "Substitute & Upload" or "Upload Original"
- Text files substituted silently (no preview friction)
- String form fields also substituted
- 30-second auto-dismiss on preview (uploads original if no response)
Preview UI (content.css):
- Centered overlay with dark theme matching existing warning UI
- Shows original → substituted pairs for each PPI found
- File format note (e.g. "PDF will be converted to plain text")
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Import parser:
- Passwords now imported from Chrome/Firefox/Bitwarden/1Password CSVs
as exact-match auto-redacted mappings (→ [REDACTED-PASSWORD-N])
- Catches passwords in any context, not just key=value patterns
Proper noun heuristic:
- Auto-detect scanner now catches capitalized words mid-sentence as
potential names, company names, or project names
- Filters against 200+ common English words, programming terms, days,
months to reduce false positives
- Added to both content.js (page world) and auto-detect.js (popup)
README:
- Documented proper noun detection with examples
- Documented bulk import with all supported formats
- Added Sync features section (auto sync, conflict resolution,
version history, connected devices)
- Added Organization/Team section with policy format
- Added tamper protection documentation
- Added Legal section with liability analysis
- Updated architecture with all new modules
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Sync encryption:
- AES-256-GCM encryption for all sync channels (browser sync, gist,
custom URL, folder sync, sync codes)
- Password with optional TOTP (RFC 6238) second factor
- Configurable auth TTL: session, 30/90/180/365 days, or never
- CryptoKey cached in IndexedDB — auth only needed when cache expires
AND new data exists (lastModified check runs before auth prompt)
- WebAuthn (biometric/PIN) as low-friction re-authentication gate
- Full options UI for setup, password change, and inline auth prompt
Smart reveal:
- Track which substitute values were actually sent outbound per session
- Reveal mode only replaces values that were genuinely substituted,
preventing false positives (e.g. AI using the word "user" won't be
replaced with a real username that maps to "user")
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
PPI + button fix:
- After adding a mapping, the real PPI value is now immediately replaced
with the fake value in the input/contenteditable element via the new
replaceInInput() helper (handles both <textarea>/<input> and
contenteditable divs by walking text nodes).
- Triggers a re-scan 150ms later so the pre-send warning updates or
dismisses itself if no more PPI remains.
Auto-sync folder (File System Access API):
- User picks a folder once per browser via "Choose Sync Folder".
Any settings/mappings/identity change writes silent-send-sync.json
to that folder automatically (via api.storage.onChanged listener).
- On options page open and every time the page regains focus, the file
is read back; if its lastModified is newer than local data, settings
are imported immediately and the UI refreshes.
- File handle is stored in IndexedDB (ss_sync_handles) so it persists
across browser sessions without repeated permission prompts.
- Pick the SAME folder in each browser (or a synced cloud folder for
cross-computer sync) — fully automatic after that, no copy-paste.
- sync.js gains saveSyncDirHandle / loadSyncDirHandle / clearSyncDirHandle
helpers backed by IndexedDB.
https://claude.ai/code/session_01TKpSR9M8JgHLXCp5CeDsQP
Bug fixes:
- Date (possible DOB) pattern now requires context words (born, birthday,
dob, etc.) before firing — prevents spurious warnings on page-load API
calls that happen to contain ISO dates in conversation history.
- Highlight regex now uses word boundaries (\b) so short substitute values
(e.g. "aud") no longer match inside unrelated words like "Claude".
- Both TreeWalkers in content.js now skip the extension's own UI elements
(.ss-autodetect-warning, .ss-presend-warning, .ss-reveal-badge) to
prevent the highlight API from marking text in the extension's banners.
Settings sync:
- New src/lib/sync.js: exportSyncCode / importSyncCode (base64 JSON) for
manual copy-paste across any browser combination. Newest lastModified
timestamp wins; force flag available to override.
- browser.storage.sync support: when "Browser account sync" is enabled the
extension automatically pushes/pulls via Firefox Sync or Chrome account,
chunked to stay within per-item quota limits.
- storage.js now writes ss_lastModified on every save so conflict resolution
has an accurate timestamp.
- service-worker.js listens for both local and sync storage changes to keep
all copies in sync.
- New "Sync Between Browsers" section in options.html with Generate/Copy/
Import Sync Code UI and the browser sync toggle.
https://claude.ai/code/session_01TKpSR9M8JgHLXCp5CeDsQP
Detected PPI is now auto-redacted in the fetch hook using
RFC/standard reserved values — not just warned about:
- IPs → 192.0.2.1 (RFC 5737 TEST-NET-1, never routed)
- MACs → 00:00:00:00:00:00
- Addresses → 123 Example Street, Anytown, ST 00000
- GPS → 0.000000,0.000000 (Gulf of Guinea)
- Dates → 01/01/1970 (Unix epoch)
- EINs → 00-0000000 (impossible prefix)
- Paths → /home/user
- Git → example org
These are obviously fake and guaranteed not to be real data,
unlike random values which could be confused with actual PPI.
New Options toggle: "Auto-redact detected PPI on send" (on by
default). When on, PPI is caught in the fetch hook even if the
user hits Enter immediately after pasting. Warning banner now
says "Auto-redacted with standard placeholders" instead of
"These were sent as-is."
https://claude.ai/code/session_01Dvgwe7XMoSxnWXkih8p1Cw
Like spellcheck for privacy. Scans text as you type and paste
into chat inputs (debounced 800ms). Shows a dark floating warning
panel listing detected PPI BEFORE you hit Enter.
Each detected item has a green [+] button that instantly:
1. Generates a plausible fake value (random IP, fake address, etc.)
2. Adds it as a mapping to storage
3. Shows a checkmark to confirm
The warning disappears when you clear the text or when all
detected items have been addressed.
Three new Options toggles:
- Auto-detect unconfigured PPI (on by default)
- Offer to auto-add detected PPI (on by default)
Also adds a storage bridge (postMessage) so the page-world
content script can read/write chrome.storage through the
injector.
https://claude.ai/code/session_01Dvgwe7XMoSxnWXkih8p1Cw
Scans outbound messages AFTER all substitutions for personal data
the user forgot to configure:
- Private/public IP addresses (skips 127.0.0.1, 8.8.8.8, etc.)
- MAC addresses
- Street addresses ("123 Main St")
- GPS coordinates
- Dates (possible DOBs)
- EIN/tax IDs
- Home directory paths not caught by smart patterns
- Shell prompts (user@host)
- Git remotes (reveals username/org)
- Environment variable assignments (HOME=, USER=, etc.)
Shows a floating dark warning banner (top-right, auto-dismisses
after 15s) listing each detected item with its type, value, and
hint. Skips values already in the user's identity config.
Also shows PPI warnings in the popup Test tab and adds toggle
in Options to disable auto-detect.
https://claude.ai/code/session_01Dvgwe7XMoSxnWXkih8p1Cw
Two visual modes using the Highlight API (zero DOM modification):
- Yellow highlight: fake/substituted values the AI received.
Shows automatically in all AI responses, even without reveal mode.
"The AI sees these yellow values, not your real data."
- Terminal style (black bg, green text): your real data shown
in reveal mode. "This is YOUR data — the AI never saw this."
Uses CSS.highlights with ::highlight() pseudo-elements — no
spans, no DOM changes, no copy/paste artifacts. Falls back to
CSS class-based styling if Highlight API is unavailable.
Highlights are debounced (500ms) to handle streaming responses
without excessive reflows.
https://claude.ai/code/session_01Dvgwe7XMoSxnWXkih8p1Cw
Revealed values are now wrapped in <span class="ss-revealed">
with a blue underline and subtle blue background, making it
obvious which parts are your real data vs what the AI said.
Hover shows tooltip "Substituted value: [fake]" so you can see
what the AI actually received. Un-reveal properly removes the
spans and normalizes text nodes.
https://claude.ai/code/session_01Dvgwe7XMoSxnWXkih8p1Cw
Storage issue: Chromium/Brave keeps storage.local data even when
an unpacked extension is removed and reloaded. Added "Reset
Everything" button in Options → Danger Zone that clears all data
(double-confirm to prevent accidents).
Reveal mode: Now re-runs revealAllResponses() every 2 seconds
while active to catch streamed content and dynamically loaded
responses. Adds console logging for reveal toggle state changes
to aid debugging. Cleans up interval when reveal mode is turned off.
https://claude.ai/code/session_01Dvgwe7XMoSxnWXkih8p1Cw
Interception is now completely inactive until the user configures
at least one identity field or explicit mapping. Before that:
- Icon shows gray (unconfigured)
- No fetch/XHR hooks fire
- First-run banner tells user to set up
Custom domains: clicking "Add Domain" in Options now triggers
the browser's native permission prompt via permissions.request().
No more manual chrome://extensions site access step.
Icon states are now:
- Gray = unconfigured (nothing will happen)
- Black = active and protecting
- Blue = reveal mode on
- Red = manually disabled
https://claude.ai/code/session_01Dvgwe7XMoSxnWXkih8p1Cw
Adds a secret scanning layer that runs after identity/explicit
substitutions. Catches secrets the user didn't configure:
- API keys: OpenAI (sk-), Anthropic (sk-ant-), Google (AIza),
AWS (AKIA), GitHub (ghp_), GitLab (glpat-), Slack (xox),
Stripe (sk_live/test), SendGrid (SG.)
- Auth: Bearer tokens, key=value assignments (password=, secret=,
api_key=, token=), private key blocks (-----BEGIN PRIVATE KEY-----)
- Connection strings: mongodb://, postgres://, mysql:// with creds
- PII: SSN (xxx-xx-xxxx), credit card numbers (Visa/MC/Amex/Discover)
Redacted values shown in red in the Test tab. Secrets are truncated
in the activity log (first 8 chars + "...") to avoid logging the
full secret. Enabled by default, toggle in Options.
https://claude.ai/code/session_01Dvgwe7XMoSxnWXkih8p1Cw
Replace URL-pattern and body-shape matching with universal approach:
Outbound (substitution):
- Hooks ALL POST/PUT/PATCH fetch and XHR requests, not just known
API endpoints. Skips static assets and analytics.
- Deep-walks any JSON structure recursively to find and substitute
all string values. Skips metadata keys (model, id, token, etc.).
- Falls back to raw string substitution for non-JSON bodies.
Inbound (reveal):
- Walks ALL text nodes in document.body, not just specific CSS
selectors. Skips SCRIPT, STYLE, INPUT, TEXTAREA tags.
- Handles streaming by observing characterData mutations globally.
This makes Silent Send survive any API restructuring — the only
thing that could break it is a site encrypting request bodies in
JS before fetch, which would also break their own dev tools.
Performance: reveal pairs are cached and only rebuilt on config
change. Deep walk skips known non-content keys to avoid touching
auth tokens or request metadata.
https://claude.ai/code/session_01Dvgwe7XMoSxnWXkih8p1Cw
Reveal mode (eye icon) now works as intended:
- Toggle ON: all existing responses on the page get fake→real
substitution applied immediately (paths, names, emails, etc.)
- Streaming responses are revealed in real-time as they arrive
- Toggle OFF: original text is restored from saved state
- Covers code blocks, artifacts, pre tags, and all response
containers across all supported services
- Blue floating badge shows "Reveal Mode — showing real data"
when active so user knows what they're seeing
The workflow is now: type /home/jsmith/... → Claude sees
/home/ademo/... → Claude responds with /home/ademo/... →
reveal mode shows /home/jsmith/... → user copies real path.
Also adds privacy note in popup footer: data stays in local
browser storage, no servers, no tracking, no analytics.
https://claude.ai/code/session_01Dvgwe7XMoSxnWXkih8p1Cw
Claude