STA-012 Threads: When a deep link becomes a persistent crash loop

Wednesday, August 26, 2026

STA-012 Threads: When a deep link becomes a persistent crash loop

Structured Text Amplification Vector STA 012

This post documents STA-012, a persistent denial-of-service vector in Meta's Threads app for Android. An oversized payload delivered via a deep link can contaminate the app's SavedState, causing a permanent crash loop that requires clearing app data to recover.

⚠️ Severity: STA-012 is a persistent DoS vector. A single click on a crafted deep link can make Threads permanently unusable until the user clears app data. No special permissions or privileges are required.


1. Summary

STA-012 describes a persistent denial-of-service condition in Meta's Threads app for Android. An attacker can craft a deep link with an oversized q parameter that, when opened, contaminates the app's SavedState. The app subsequently enters a permanent crash loop that only resolves by clearing the app data.

The vector is triggered when the user clicks a crafted deep link. Threads stores the search query in Fragment arguments, which are then serialized into a Bundle as part of the SavedState. When the Bundle size exceeds the Binder transaction limit (approximately 1 MB), a TransactionTooLargeException is thrown during state restoration. The app crashes, and the same oversized state is re-persisted, causing a crash on every subsequent launch.

Type: Persistent Denial of Service (DoS)
Class: Resource Exhaustion / State Persistence Boundary
CWE: CWE-400 — Uncontrolled Resource Consumption; CWE-770 — Allocation of Resources Without Limits
CVSS v3.1 estimated: 7.1 (High)
Vector: AV:N/AC:L/PR:N/UI:R/S:U/C:N/I:N/A:H


2. Attack chain

The observed chain can be represented as:

Attacker crafts deep link
       ↓
https://www.threads.net/search?q=[OVERSIZED_PAYLOAD]
       ↓
User clicks the link
       ↓
Threads opens and handles the deep link
       ↓
Search query stored in Fragment arguments
       ↓
Fragment state saved as SavedState
       ↓
Bundle serialization (FragmentManager)
       ↓
Bundle size exceeds Binder limit
       ↓
TransactionTooLargeException
       ↓
Threads crashes
       ↓
State re-persisted to disk
       ↓
Threads crashes on every launch attempt
       ↓
PERSISTENT CRASH LOOP
       ↓
Recovery: clear app data or uninstall/reinstall

The fundamental characteristic of STA-012 is that the payload is delivered through a standard deep link that the app is designed to handle. The user does nothing unusual — just clicks a link that appears legitimate.


3. Technical details

3.1. Deep link structure

The crafted deep link targets Threads' search functionality:

https://www.threads.net/search?q=[PAYLOAD]

The q parameter contains the oversized payload, which is accepted by the app and stored in Fragment arguments as part of the navigation state.

3.2. State amplification

The payload is amplified through the same mechanism documented in STA-005 (WhatsApp):

Input (payload in deep link)
       ↓
Fragment arguments (SavedState)
       ↓
Bundle serialization
       ↓
Nested fragments add metadata overhead
       ↓
Total Bundle size exceeds Binder limit
       ↓
TransactionTooLargeException

From the structural amplification paper, observed Bundle component sizes in similar Class A vectors:

androidx.lifecycle.BundlableSavedStateRegistry.key: 1,661,088 bytes
android:support:fragments: 1,645,336 bytes
childFragmentManager: 1,636,412 bytes
childFragmentManager: 1,614,072 bytes
childFragmentManager: 1,603,852 bytes
registryState: 479,948 bytes
search_query: 80,484 bytes  ← input data

Key observation: The search query (80,484 bytes) is amplified to a total Bundle size of 1.66 MB — exceeding the Binder limit by 58%.

3.3. Amplification factor

Based on the observed Bundle sizes, the amplification factor for the search_query component is:

Input: 80,484 bytes
Total Bundle: 1,661,088 bytes
Amplification factor: ×20.6

This is consistent with the amplification factor observed in STA-005 (WhatsApp).


4. Stack trace

The following stack trace was captured during a Threads crash caused by an oversized deep link payload. The excerpt is abbreviated; irrelevant frames and build-specific details have been omitted.

android.os.TransactionTooLargeException: data parcel size 1661088 bytes
    at android.os.BinderProxy.transactNative(Native Method)
    at android.os.BinderProxy.transact(BinderProxy.java)
    at android.app.ActivityManagerProxy.activityStopped(ActivityManagerProxy.java)
    at android.app.ActivityThread.handleStopActivity(ActivityThread.java)
    at android.app.ActivityThread$H.handleMessage(ActivityThread.java)
    at android.os.Handler.dispatchMessage(Handler.java)
    at android.os.Looper.loop(Looper.java)
    at android.app.ActivityThread.main(ActivityThread.java)
    at java.lang.reflect.Method.invoke(Native Method)
    at com.android.internal.os.RuntimeInit$MethodAndArgsCaller.run(RuntimeInit.java)
    at com.android.internal.os.ZygoteInit.main(ZygoteInit.java)

Key metric: Parcel size 1,661,088 bytes — exceeds the Binder transaction limit (1,048,576 bytes) by approximately 58%.


5. Variants

STA-012e — Threads via WhatsApp

An attacker can deliver the payload through a WhatsApp message containing a WebView or HTML that triggers the Threads deep link:

WhatsApp message
       ↓
User clicks link/HTML
       ↓
WebView opens
       ↓
Redirect to Threads deep link
       ↓
https://www.threads.net/search?q=[PAYLOAD]
       ↓
Threads opens → crash loop

This expands the attack surface, as the payload can be delivered through messaging platforms rather than requiring direct navigation to a web page.


6. Persistence

Persistence: YES

The contaminated state is written to disk. The crash loop persists across:

  • App restarts
  • Device reboots
  • App updates (if the state is preserved)

Recovery requires:

  • Clearing the app data (via adb pm clear com.threads or app settings)
  • Uninstalling and reinstalling the app (loses all data)

7. Affected components

Primary

  • androidx.fragment.app.FragmentManager — state restoration
  • androidx.lifecycle.SavedStateRegistry — state persistence
  • androidx.navigation.NavController — navigation state
  • android.os.Bundle — serialization container
  • android.os.Parcel — Binder serialization

Secondary

  • android.app.ActivityThread — lifecycle handling
  • android.app.ActivityManagerProxy — Binder transaction

8. Vendor status

Vendor Status
Meta (Threads) ❌ Reported in March 2026 — no response as of August 2026
Google (Android VRP) ✅ A-477279924 — $250 reward, open triage
Google (AndroidX) ✅ savedstate 1.5.0 (May 2026) decouples SavedState from Binder — but does not fix FragmentManager or apps using older patterns

Note: Updating to androidx.savedstate 1.5.0 does not protect against STA-012. Threads must update their own state handling, and FragmentManager.restoreAllState() itself remains unpatched.


9. Why this matters

STA-012 is not just a crash. It is a permanent denial of service with significant implications:

  • For individual users: They lose access to the app until they clear data, losing all conversations and settings
  • For the platform: This is a single point of failure in the state restoration mechanism
  • For Meta: A single malicious link can make Threads unusable for any user who clicks it

The vector is particularly concerning because:

  • The payload is delivered through a standard deep link that the app is designed to handle
  • No special permissions are required
  • The user does nothing unusual — just clicks a link
  • The impact is persistent and requires clearing data to recover

10. Relationship to other STA vectors

Vector Relationship
STA-005 Same mechanism (text input → SavedState → Bundle → Binder → crash loop). WhatsApp has a similar amplification factor (×20.6).
STA-015-DL Same amplification mechanism, escalated to SystemUI via TaskPersister.
STA-022 Same mechanism (DuckDuckGo Fragment args → Parcel → Binder).
STA-028 UTF-16 encoding amplification is a contributing factor to the measured ×20.6 amplification.

11. Recommended mitigation

11.1. Application-level (Threads)

  • Validate the size of deep link parameters before storing them in Fragment arguments
  • Truncate search queries to a safe limit (e.g., 8 KB)
  • Catch TransactionTooLargeException in state restoration and fall back to clean start
// Recommended approach for deep link handling
Uri data = getIntent().getData();
String query = data.getQueryParameter("q");
if (query != null && query.length() > MAX_SAFE_QUERY_LENGTH) {
    query = query.substring(0, MAX_SAFE_QUERY_LENGTH);
    Log.w(TAG, "Search query truncated to safe length");
}

11.2. Framework-level (Android)

  • FragmentManager.restoreAllState() should catch TransactionTooLargeException and discard oversized state
  • Bundle should provide a size estimation method before serialization
  • AndroidX Navigation should validate argument sizes before persisting state

12. Research status

Field Value
Vector STA-012
First formal communication 20 January 2026
Researcher Manuel García Peña (Lostmon)
Nature Independent research
Platform Android
Impact Persistent DoS
Interaction required Yes (one click)
Privileges None
Tier A (Confirmed — full stack trace + exception)

STA-012 is part of the broader Structured Text Amplification (STA) research, which studies a recurring pattern of resource exhaustion produced when input data crosses serialization, transformation, IPC, or persistence boundaries without sufficiently early resource limits.


13. Conclusion

STA-012 demonstrates how a standard deep link can become a persistent denial-of-service vector when the app fails to validate the size of input parameters before storing them in SavedState.

The amplification factor (×20.6) is consistent with other Class A vectors, confirming that the problem is architectural rather than specific to a single app:

Deep link payload
 → Fragment arguments
 → SavedState
 → Bundle
 → Parcel
 → Binder
 → TransactionTooLargeException
 → Persistent crash loop

The most robust mitigation is to validate input size before storing it in state, complemented by fallback mechanisms that prevent an oversize condition from becoming a persistent crash loop.


Complete whitepaper: Resilience Gaps in Android IPC, SavedState and Text Layout — v6 (August 2026)


📌 About this series
This post is part of a series documenting the 32 vectors of Structured Text Amplification (STA).

Published:
STA-017 — Cross-Engine ANR
STA-015-DL — Google Drive → SystemUI

STA-006 / 007 Translate: When translating text crashes Google Apps

STA-006 / 007 Translate: When translating text crashes Google Apps

Structured Text Amplification — Vectors 006 & 007

This post documents STA-006 and STA-007, two denial-of-service vectors in Google's Translate feature. An oversized text payload, when selected and sent to Translate via Intent.ACTION_TRANSLATE, can exceed the Binder transaction limit, causing a TransactionTooLargeException and a permanent crash loop in some cases.

⚠️ Severity: STA-006 and STA-007 are persistent DoS vectors in some oems implementations. A single oversized text selection can make Google App or Google Drive permanently unusable until the user clears app data. No special permissions or privileges are required.


1. Overview

STA-006 and STA-007 describe a persistent denial-of-service condition in Google's Translate feature, accessible from multiple Google apps. An attacker can craft or deliver an oversized text payload that, when selected and translated, contaminates the app's state and causes a permanent crash loop in oems.

The vector is triggered when the user selects a large text (visible or hidden) and invokes the Translate action. The app constructs an Intent.ACTION_TRANSLATE containing the selected text. During serialization of the Intent for transfer via Binder, the data size may exceed the Binder transaction limit (approximately 1 MiB). This can result in an unhandled TransactionTooLargeException, causing an immediate app crash.

Type: semi-persistent Denial of Service (DoS)
Class: Resource Exhaustion / State Persistence Boundary
CWE: CWE-400 — Uncontrolled Resource Consumption; CWE-770 — Allocation of Resources Without Limits
CVSS v3.1 estimated: 6.5 (Medium-High)
Vector: AV:N/AC:L/PR:N/UI:R/S:U/C:N/I:N/A:H


2. Attack chain

The observed chain can be represented as:

Attacker crafts/places oversized text payload
       ↓
User selects the text (or selects all in a PDF)
       ↓
User invokes Translate action (context menu or app button)
       ↓
Google App / Google Drive constructs ACTION_TRANSLATE Intent
       ↓
Selected text stored in Intent extra (EXTRA_TEXT)
       ↓
Intent serialization (Parcel)
       ↓
Parcel size exceeds Binder limit (1 MB)
       ↓
TransactionTooLargeException
       ↓
App crashes immediately
       ↓
State may be re-persisted (in Google Drive)
       ↓
App crashes on every launch attempt (persistent crash loop)
       ↓
Recovery: clear app data or uninstall/reinstall

The fundamental characteristic of STA-006/007 is that the STA payload is delivered through a standard user action (Translate) that the app is designed to handle. The user does nothing unusual — just selects text and translates it.


3. Vector details

3.1. STA-006 — Google App (Select text → Translate)

The vector is triggered through the Google App's selection menu:

  • The user selects a large text (e.g., from a web page, document, or any text field).
  • The user taps the "Translate" action from the context menu.
  • Google App creates an ACTION_TRANSLATE Intent with the selected text as EXTRA_TEXT.
  • During serialization of the Intent, the Bundle size exceeds the Binder transaction limit.
  • TransactionTooLargeException is thrown and not caught.
  • Google App crashes immediately.

Entry point: android.app.Activity.startActivityForResult()Instrumentation.execStartActivity() → Binder transaction → TransactionTooLargeException

3.2. STA-007 — Google Drive (PDF → Select All → Translate)

The vector is triggered through Google Drive's PDF viewer:

  • User opens a PDF containing invisible or oversized text (e.g., a PDF with a large hidden payload).
  • User selects all text (Select All) in the PDF viewer.
  • User taps the "Translate" action from the selection menu.
  • Google Drive creates an ACTION_TRANSLATE Intent with the selected text.
  • During serialization of the Intent, the Bundle size exceeds the Binder transaction limit.
  • TransactionTooLargeException is thrown and not caught.
  • Google Drive crashes immediately.
  • If the state is re-persisted, the app enters a permanent crash loop.

Entry point: PDF viewer selection menu → Intent.ACTION_TRANSLATE → Binder transaction → TransactionTooLargeException


4. Stack trace and Bundle analysis

The following stack trace and Bundle statistics were captured from a production device (Xiaomi Redmi Note 14 5G, HyperOS 3.0, Android 16) during a crash of the Google App. The crash occurred after selecting an oversized text payload, invoking Translate, and subsequently focusing the search bar, which triggered the state serialization and Binder transaction.

4.1. Exception

java.lang.RuntimeException: android.os.TransactionTooLargeException:
data parcel size 2454260 bytes
    at android.app.servertransaction.PendingTransactionActions$StopInfo.run(PendingTransactionActions.java:146)
    at android.os.Handler.handleCallback(Handler.java:1029)
    at android.os.Handler.dispatchMessage(Handler.java:107)
    at android.os.Looper.loopOnce(Looper.java:274)
    at android.os.Looper.loop(Looper.java:369)
    at android.app.ActivityThread.main(ActivityThread.java:10090)
    at java.lang.reflect.Method.invoke(Native Method)
    at com.android.internal.os.RuntimeInit$MethodAndArgsCaller.run(RuntimeInit.java:616)
    at com.android.internal.os.ZygoteInit.main(ZygoteInit.java:1137)
Caused by:
android.os.TransactionTooLargeException:
data parcel size 2454260 bytes
    at android.os.BinderProxy.transactNative(Native Method)
    at android.os.BinderProxy.transact(BinderProxy.java:736)
    at android.app.IActivityClientController$Stub$Proxy.activityStopped(IActivityClientController.java:1546)
    at android.app.ActivityClient.activityStopped(ActivityClient.java:106)
    at android.app.servertransaction.PendingTransactionActions$StopInfo.run(PendingTransactionActions.java:135)

4.2. Bundle statistics

The Bundle that triggered the exception contained the following components:

Bundle stats:
    androidx.lifecycle.BundlableSavedStateRegistry.key [size=2452080]
    androidx.lifecycle.internal.SavedStateHandlesProvider [size=1127400]
    tt_activity_account_retained:6 [size=1125236]
    SearchSession:1 [size=1125012]
    androidx.lifecycle.BundlableSavedStateRegistry.key [size=1124884]
    androidx.lifecycle.internal.SavedStateHandlesProvider [size=1124752]
    androidx.lifecycle.ViewModelProvider.DefaultKey:... [size=1124452]
    BrowserSession:2 [size=963088]
    androidx.lifecycle.BundlableSavedStateRegistry.key [size=962960]
    androidx.lifecycle.internal.SavedStateHandlesProvider [size=962828]
    webx_browser [size=962464]
    webx_window [size=962416]
    web_view_url [size=962288]
    search_query_key [size=161080]
    search_query [size=161028]
    protoparsers [size=160976]
    android:support:fragments [size=1312836]
    fragment_e32da9b3... [size=1312412]
    arguments [size=161152]
    TIKTOK_FRAGMENT_ARGUMENT [size=161008]
    childFragmentManager [size=1312836]
    android-support-nav:controller:backStack [size=326252]
    registryState [size=168492]
    SEARCH_QUERY_STATE [size=160976]
    ... (additional nested fragment states)

4.3. Key metrics

Component Size
Input (search_query) ~161,028 bytes
childFragmentManager 1,312,836 bytes
SavedStateHandlesProvider 1,127,400 bytes
SearchSession / BrowserSession ~1,125,000 bytes each
Total Bundle size 2,454,260 bytes
Binder transaction limit 1,048,576 bytes
Excess +1,405,684 bytes (+134%)
Amplification factor ×15.2

Key observation: The input payload (~161 KB) is amplified to a total Bundle size of 2.45 MB through the serialization of nested Fragment states, SavedState handles, and session data. The amplification factor of ×15.2 is consistent with the amplification patterns documented in other Class A vectors (STA-005, ×20.6; STA-012, ×20.6).


5. Affected components

Primary

  • android.content.IntentACTION_TRANSLATE intent extra serialization
  • android.os.Bundle — serialization container
  • android.os.Parcel — Binder serialization
  • android.app.ActivityThread — lifecycle handling
  • androidx.fragment.app.FragmentManager — fragment state serialization
  • androidx.lifecycle.SavedStateHandlesProvider — SavedState management

Secondary

  • android.app.ActivityManagerProxy — Binder transaction
  • Google App (com.google.android.googlequicksearchbox) — Translate feature
  • Google Drive (com.google.android.apps.docs) — PDF viewer + Translate feature

6. Persistence

Persistence: YES (STA-007), Conditional (STA-006)

  • STA-006 (Google App): The app may crash immediately, but the state is not always re-persisted. Recovery typically involves reopening the app.
  • STA-007 (Google Drive): The PDF viewer state is saved, and the oversized payload may be re-persisted, causing a crash on every launch.

Recovery typically requires:

  • Clearing the app data (via adb pm clear com.google.android.googlequicksearchbox or adb pm clear com.google.android.apps.docs)
  • Uninstalling and reinstalling the app (loses all data)

7. Why this matters

STA-006 and STA-007 are not just crashes. They are semi-persistent denial-of-service vectors with significant implications:

  • For individual users: They lose access to the app until they clear data, losing all settings and potentially documents.
  • For Google: A single malicious document can make Google Drive unusable for any user who opens it and attempts to translate.
  • For the platform: This is a single point of failure in the Intent serialization and Binder transaction mechanism.

The vector is particularly concerning because:

  • The payload can be delivered through a standard PDF document that the app is designed to handle.
  • No special permissions are required.
  • The user does nothing unusual — just selects text and translates it.
  • The impact is persistent in some OEMS and requires clearing data to recover.

8. Relationship to other STA vectors

Vector Relationship
STA-003 Same mechanism (Share Intent → Binder → TransactionTooLargeException)
STA-005 Same mechanism (text input → SavedState → Bundle → Binder → crash loop). WhatsApp has a similar amplification factor (×20.6).
STA-012 Same mechanism (deep link → Fragment args → SavedState → Binder → crash loop)
STA-028 UTF-16 encoding amplification is a contributing factor to the measured amplification.

9. Chromium commit evidence

Chromium commit 84b615a0 (5 May 2026) added the Translate action to the Android PDF viewer's selection menu, confirming that this surface is considered relevant and that Google engineers have been working on it. The commit introduced:

84b615a07fc14988b46f0e362502ab4075216793
Author: Ryan Thomas
Date: 5 May 2026
Component: Android PDF viewer selection menu
Changes: Added Share, Web Search, Translate actions to selection menu

This demonstrates that the surface identified by STA-006/007 is being actively developed and modified, further validating the architectural relevance of the findings.


10. Recommended mitigation

10.1. Application-level (Google App / Google Drive)

  • Validate the size of selected text before constructing the ACTION_TRANSLATE Intent.
  • Truncate text to a safe limit (e.g., 8 KB) before placing it in EXTRA_TEXT.
  • Catch TransactionTooLargeException and fall back to a clean state.
// Recommended approach for Translate action
String selectedText = getSelectedText();
if (selectedText != null && selectedText.length() > MAX_SAFE_LENGTH) {
    selectedText = selectedText.substring(0, MAX_SAFE_LENGTH);
    Log.w(TAG, "Selected text truncated to safe length");
}
Intent translateIntent = new Intent(Intent.ACTION_TRANSLATE);
translateIntent.putExtra(Intent.EXTRA_TEXT, selectedText);
try {
    startActivity(translateIntent);
} catch (TransactionTooLargeException e) {
    Log.e(TAG, "Translate Intent too large", e);
    // Fallback: show error dialog or use alternative translation method
}

10.2. Framework-level (Android)

  • Intent serialization should include a size estimation before Binder transaction.
  • SystemUI should catch TransactionTooLargeException and handle it gracefully.
  • FragmentManager should limit the size of saved state before serialization.

11. Research status

Field Value
Vectors STA-006, STA-007
First formal communication 20 January 2026
Researcher Manuel García Peña (Lostmon)
Nature Independent research
Platform Android
Impact Persistent DoS
Interaction required Yes (one click / selection)
Privileges None
Tier A (Confirmed — full stack trace + exception)

STA-006/007 are part of the broader Structured Text Amplification (STA) research, which studies a recurring pattern of resource exhaustion produced when input data crosses serialization, transformation, IPC, or persistence boundaries without sufficiently early resource limits.


12. Conclusion

STA-006 and STA-007 demonstrate how a standard Translate action can become a persistent denial-of-service vector when the app fails to validate the size of selected text before constructing an Intent.

The amplification mechanism is consistent with other Class A vectors, confirming that the problem is architectural rather than specific to a single app:

Selected text (~161 KB)
 → Intent (ACTION_TRANSLATE)
 → FragmentManager serialization
 → SavedStateHandlesProvider
 → Bundle (2.45 MB)
 → Parcel
 → Binder
 → TransactionTooLargeException
 → Persistent crash loop

The most robust mitigation is to validate input size before constructing the Intent, complemented by fallback mechanisms that prevent an oversize condition from becoming a persistent crash loop.


Complete whitepaper: Resilience Gaps in Android IPC, SavedState and Text Layout — v6 (August 2026)


📌 About this series
This post is part of a series documenting the 32 vectors of Structured Text Amplification (STA).

Published:
STA-017 — Cross-Engine ANR
STA-015-DL — Google Drive → SystemUI
STA-005 — WhatsApp
STA-003 — Binder Share Intent
STA-006/007 — Translate (this post)

Coming next:
⬜ STA-012 — Threads
⬜ STA-019 — Firefox
⬜ STA-022 — DuckDuckGo

Whitepaper: Resilience Gaps in Android IPC, SavedState and Text Layout — v6


Lostmon · lostmon.blogspot.com

STA - Structured Text Amplification In Llm's

Sunday, August 23, 2026

🧩 Structured Text Amplification (STA)

A Systemic Vulnerability in LLMs. Documented from the Couch
📅 August 23, 2026 👤 Lostmon 🏷️ Research / Vulnerability / Tokenization

Structured Text Amplification (STA) is a phenomenon where a finite-length input sequence, composed of non-semantic characters and lacking structural delimiters, causes a non-linear growth in computational cost in generative AI systems.

We tested 7 different systems (DeepSeek, Grok, Gemini, Copilot, Leo, Qwen VL, and others) and all are vulnerable, though with different symptoms: reasoning loops, 19-minute thinking times, parsing errors, interface amplification, and more.

🧠 The Key: STA is not a flaw in a specific model, but a structural problem in the design of AI systems — affecting tokenization, the ingestion interface, the parser, and the reasoning mode.

⚙️ STA Pattern Used

The base pattern is a repetition of special characters without separators or semantic meaning

We tested lengths of 1,600, 10,000, 61,560, and 65,560 characters, always with UTF-8 encoding.

📊 Results by System

SystemInputMain SymptomAmplification
DeepSeek (V4-Flash)1,600 charsReasoning loop, long responses~4.4x
Grok (xAI)1,600 chars"Think" mode activated for 19 min without response
Gemini (Google)61,560 charsLong structured response, no useful data
Copilot (GitHub)61,560 charsInflated count (1,002,682) + fragmentation16.28x
Leo (Mistral)65,560 charsSyntax error: "Unterminated string"
Qwen VL 30B65,560 charsSyntax error: "Unterminated string"

🧬 Layer-by-Layer Analysis

LayerVulnerabilityAffected Systems
Input ParserDoes not escape special characters → syntax errorLeo, Qwen VL
Ingestion InterfaceConverts long text into truncated document with repetitionsCopilot
Tokenizer (BPE)Fragments special characters as individual tokensDeepSeek, Grok, Gemini
Inference EngineEnters loop without semantic structureDeepSeek, Grok
Reasoning ModeSTA prevents convergence → prolonged blockGrok

🎯 Attack Vectors Identified

  • Direct Vector: Sending the STA pattern as a message to the model (1,600 characters).
  • File Vector: Uploading the pattern in a file (61,560 characters).
  • Interface Vector: Pasting the pattern into an interface that converts it to a document (Copilot).
  • Multimodal Vector: Including the pattern in an image/text context (Qwen VL).
  • Clipboard Vector: Fragmentation and contamination of the clipboard (Copilot).

🔍 The Copilot Case: Interface Amplification

Copilot does not amplify STA by itself; rather, the interface converts the long text into an internal document (<AttachedDocument>), truncates it, and fills it with repeated blocks. The model receives that amplified document and processes it as if it were real.

Actual input: 61,560 characters Internal document: ~1,002,682 characters Amplification factor: 16.28x

This is especially serious because the user has no control over this process, and the attack can escalate without the model or the user detecting it.

💰 Estimated Economic Impact

ModelInput (chars)Approx. Cost per Attack
DeepSeek1,600$0.0014
OpenAI GPT-4 (reference)1,600$0.06
Copilot (with amplification)61,560 → 1MNot quantified, but high

If the attack is automated (10 requests/second), costs can quickly escalate to hundreds of dollars per hour.

🛡️ Technical Recommendations

  • Parser: Automatically escape non-ASCII characters and validate string termination.
  • Ingestion Interface: Do not convert long text into internal documents if it is not an explicitly uploaded file. If converted, do not truncate with repetitions.
  • Tokenizer: Add subword merging rules for common combinations of special characters and limit the number of tokens per input.
  • Inference Engine: Implement timeouts in reasoning mode and detect low-entropy patterns to respond with an error without spending tokens on inference.
  • User: Do not paste long strings of special characters into AI interfaces; use preprocessing tools that clean non-semantic characters.

📌 Conclusions

  • STA is a real and documented phenomenon affecting generative AI systems across multiple layers.
  • All tested models are vulnerable, though with different symptoms.
  • The interface layer can amplify the attack (Copilot: 16.6x).
  • STA is not a flaw in a specific model, but a structural problem in the design of AI systems.
  • Urgent action is recommended to mitigate this attack vector.

📚 References & Further Reading


🛋️ Research led from the couch with ingenuity, patience, and insatiable curiosity.
Lostmon
🧩 STA — Structured Text Amplification  ·  Version 2.0 (Technical)  ·  Published under CC BY-NC 4.0
```

STA-003 When sharing an oversized link breaks the browser (Binder Share Intent)

Friday, August 21, 2026

STA-003 Binder Share Intent TransactionTooLargeException

Structured Text Amplification — Vector 003

This post documents STA-003, a denial-of-service vector in Android browsers where an oversized URL shared via the context menu triggers a TransactionTooLargeException during Binder serialization of the ACTION_SEND Intent, causing an immediate browser crash.

⚠️ Severity: STA-003 is an interactive DoS vector. A single long-press on a crafted link followed by "Share" can crash the browser immediately. No special permissions or privileges are required.


1. Summary

STA-003 describes a denial-of-service condition in Android web browsers when an excessively large URL is shared via the context menu of a link.

The vector is triggered when the user performs a long-press on a link and selects "Share". The browser constructs an ACTION_SEND Intent containing the URL as EXTRA_TEXT. During serialization of the Intent for transfer via Binder, the data size may exceed the Binder transaction limit (approximately 1 MiB). This can result in an unhandled TransactionTooLargeException, causing an immediate browser process termination.

Type: Denial of Service (DoS)
Class: Resource Exhaustion / IPC Serialization Boundary
CWE: CWE-20 — Improper Input Validation; CWE-400 — Uncontrolled Resource Consumption
CVSS v4.0 estimated: 6.5 (Medium)


2. Attack chain

The observed chain can be represented as:

Malicious web page
       ↓
Excessively large URL
       ↓
Long-press on the link
       ↓
Context menu
       ↓
"Share"
       ↓
Intent ACTION_SEND
       ↓
EXTRA_TEXT = URL
       ↓
Intent serialization
       ↓
Binder IPC
       ↓
Parcel > transaction limit
       ↓
TransactionTooLargeException
       ↓
Browser crash
       ↓
Denial of Service

The fundamental characteristic of STA-003 is that a legitimate-looking input — a URL — acquires a disproportionate cost when crossing a serialization/IPC boundary.


3. Difference from STA-001

STA-003 must be kept separate from STA-001.

Vector Mechanism Failure
STA-001 Context menu processing Prolonged processing / freeze before crash
STA-003 ACTION_SEND Intent → Binder serialization Immediate crash during Intent serialization

STA-001 is associated with prolonged processing during context menu construction, with a noticeable freeze before the crash.

STA-003 fails during the Share operation itself:

Long URL
       ↓
ACTION_SEND
       ↓
Intent serialization
       ↓
Binder
       ↓
TransactionTooLargeException
       ↓
Immediate crash

4. Required conditions

The attack requires:

  • An Android browser that exposes the Share operation for the affected link.
  • A page containing a link with a sufficiently large URL to exceed the effective Binder transaction limit.
  • The victim performs a long-press on the link.
  • The victim selects "Share".

Not required:

  • Special Android permissions
  • Local access to the device
  • Application privileges
  • Prior code execution on the device

The URL can be distributed via any channel capable of delivering a link, including web pages, messaging, email, social networks, QR codes, or other URL distribution mechanisms.


5. Browser-specific observations

A relevant observation during the investigation is that the availability of the vector depends on how each browser handles excessively large URLs.

Browser Behaviour Mitigation level
Google Chrome Does not present Share option for extremely large URLs Application-level
Microsoft Edge Equivalent mitigation behaviour Application-level
Opera Browser Share operation available → TransactionTooLargeException → crash No mitigation

This difference is important because it demonstrates two levels of mitigation:

Application-level mitigation:
    Avoid generating/sending an excessively large Intent

Framework-level:
    Accept an oversized operation
    ↓
    TransactionTooLargeException
    ↓
    Safe fallback / degradation

The existence of application-level mitigations does not eliminate the underlying condition in the IPC mechanism.


6. Stack trace and evidence

The following stack trace was captured during a Share operation with an oversized URL. The excerpt is abbreviated; irrelevant frames and build-specific details have been omitted.

Share Intent crash

java.lang.RuntimeException: android.os.TransactionTooLargeException:
data parcel size 1662976 bytes
at android.app.ActivityClient.activityStopped(ActivityClient.java:101)
at android.app.servertransaction.PendingTransactionActions$StopInfo.run()
at android.app.servertransaction.PendingTransactionActions$StopInfo.run(...)
at android.os.Handler.dispatchMessage(Handler.java)
at android.os.Looper.loop(Looper.java)
at android.app.ActivityThread.main(ActivityThread.java)

Key metric: Parcel size 1,662,976 bytes — exceeds the Binder transaction limit (1,048,576 bytes) by approximately 58%. The crash is immediate, with no observable delay.

Vulnerable code pattern

The following pattern is present in many browsers and apps that implement sharing functionality:

// Browser code (simplified)
Intent shareIntent = new Intent(Intent.ACTION_SEND);
shareIntent.setType("text/plain");
shareIntent.putExtra(Intent.EXTRA_TEXT, url); // ← No size check!
startActivity(Intent.createChooser(shareIntent, "Share via"));

The Android framework provides no guardrails for Intent.putExtra() size, and no automatic truncation or validation is performed before the Intent is serialized for Binder transmission.


7. Impact

The observed impact is a denial of service of the browser:

  • Immediate browser crash during the Share operation
  • Loss of browsing session
  • Interruption of user activity
  • Potential repetition of the crash if the operation is re-executed

Unlike STA-001 (which may show a 10-17 second freeze before crash), STA-003 crashes immediately during Intent serialization.

Any application implementing share functionality is potentially vulnerable:

  • Web browsers (all vendors)
  • Email clients
  • Social media apps
  • Messaging apps
  • Document viewers
  • File managers
  • Any app with a "Share" button for text/URLs

Estimated affected apps: 10,000+ on Play Store


8. Relationship to Structured Text Amplification

STA-003 belongs to the class of vectors where a legitimate-looking data item crosses a transformation or serialization boundary without a sufficiently early limit.

The pattern can be expressed as:

Small / legitimate input
        ↓
Transformation / serialization
        ↓
Structured representation
        ↓
IPC boundary
        ↓
Resource limit exceeded
        ↓
Failure / DoS

In STA-003:

URL
 ↓
String / EXTRA_TEXT
 ↓
Intent
 ↓
Parcel
 ↓
Binder
 ↓
TransactionTooLargeException
 ↓
Browser crash

The vector therefore does not depend solely on the logical size of the URL. The relevant cost is the size of the representation that must ultimately cross the IPC boundary.


9. Proof of Concept

The PoC consists of an HTML page containing a link with an extremely large URL:

<a href="https://example.com/[STA Pattern]">
    Test link
</a>

During the investigation, character patterns were used to produce an especially large representation after encoding/serialization transformations. The full payload is not included in this public documentation to reduce the potential for abuse.

Conceptual reproduction:

  1. Open an affected Android browser (e.g., Opera).
  2. Access a page containing the prepared link.
  3. Long-press the link.
  4. Select "Share".
  5. Observe the immediate browser crash.
  6. Logs may show TransactionTooLargeException with a parcel size exceeding 1 MB.

10. Evidence and background

The research relates STA-003 to the following background:

  • Chromium Issue 40879254 (2022): Issue related to sharing an excessively large URL for Binder.
  • Mozilla Bugzilla #1802594: Incidents related to TransactionTooLargeException / DeadSystemException in Firefox.
  • Chromium: Changes intended to truncate visible URLs and reduce the size of certain data before subsequent operations.
  • Android TransactionTooLargeException: Official documentation of the failure mechanism associated with excessively large Binder transactions.
  • Android Intent.ACTION_SEND: Surface used to transfer content that subsequently crosses IPC.

These references constitute independent background that helps establish that the attack surface is not limited to a specific browser implementation.

A relevant observation during the investigationis that Chrome and Edge hide the Share option for extremely large URLs, while Opera does not. This suggests that Chrome's behaviour is a deliberate application-level mitigation rather than a framework-level protection. The difference is important: it demonstrates that application-level mitigations can prevent the crash, but they do not address the underlying framework condition. A malicious payload delivered through an app that lacks such a mitigation (like Opera) still triggers the TransactionTooLargeException and crash. This aligns with Google's documented position that TransactionTooLargeException is a framework constraint, not a security vulnerability. However, the availability of application-level mitigations does not eliminate the underlying risk for apps that do not implement them.


11. Proposed mitigation

11.1. Application-level mitigation

The browser should check the URL size before constructing the Intent that will be transferred via Binder.

URL
 ↓
Size validation
 ↓
Does it exceed the limit?
 ├── Yes → truncate / reject / safe alternative
 └── No → build ACTION_SEND

This is the preferred defense because it prevents the oversized object from reaching the IPC boundary. Browsers that already hide the Share option for excessively large URLs (Chrome, Edge) provide an example of this approach.

11.2. Framework-level mitigation

Android could provide additional protection mechanisms so that a transaction exceeding the limit does not necessarily result in an unrecoverable crash of the consuming process.

Oversized transaction
        ↓
Detect before / during IPC
        ↓
Controlled failure
        ↓
Fallback
        ↓
Application remains operational

Late detection of a size condition should not automatically become a process termination condition when a safe alternative exists.

11.3. SavedState / FragmentManager (related surface)

In related SavedState surfaces, the research additionally proposes controlled fallback against errors occurring during restoration of excessively large state.

try {
    restoreStateInternal(state);
} catch (TransactionTooLargeException e) {
    Log.e(TAG, "SavedState restore failed. Restarting without state.", e);
}

The goal would be to degrade to a clean state when it is safe to do so, rather than propagating the exception to cause a crash.


12. Classification

STA-003 can be classified as:

Binder Serialization Boundary Resource Exhaustion

Within the STA model:

"Uncontrolled amplification / expansion across a serialization and IPC boundary leading to resource exhaustion."

The vector also has an important characteristic: the initial data can be completely valid from a semantic point of view — a URL — and become dangerous solely due to its size and the cost of transporting it between components.


13. Scope of the claim

STA-003 does not demonstrate that Binder is inherently vulnerable or that all Android browsers are exploitable.

The research demonstrates a more specific condition:

"When an application allows excessively large data to reach an ACTION_SEND operation and subsequently cross Binder without sufficient prior validation, the transaction size limit can become a denial-of-service mechanism."

Successful exploitation depends on the specific browser implementation and whether it incorporates validation, truncation, or fallback before constructing or sending the Intent.


14. Research status

Field Value
Vector STA-003
First formal communication 20 January 2026
Researcher Manuel García Peña (Lostmon)
Nature Independent research
Platform Android
Primary surface Browser → Intent.ACTION_SEND → Binder
Impact DoS
Interaction required Yes
Privileges None
Special permissions None

STA-003 is part of the broader Structured Text Amplification (STA) research, which studies a recurring pattern of resource exhaustion produced when input data crosses serialization, transformation, IPC, or persistence boundaries without sufficiently early resource limits.


15. Conclusion

STA-003 demonstrates how an ordinary operation — sharing a link — can become a denial-of-service condition when an oversized input crosses multiple representation layers:

URL
 → Intent
 → Parcel
 → Binder
 → transaction limit
 → TransactionTooLargeException
 → crash

The most robust mitigation is to measure and limit the size before reaching the amplification or IPC boundary, complemented by fallback mechanisms that prevent an oversize condition from unnecessarily becoming a process crash.

In the context of STA, this vector constitutes an example of how a legitimate-looking structured input can become a resource exhaustion condition when crossing a serialization boundary.


Complete whitepaper: Resilience Gaps in Android IPC, SavedState and Text Layout — v6 (August 2026)


📌 About this series
This post is part of a series documenting the 32 vectors of Structured Text Amplification (STA).

Published:
STA-005 — WhatsApp
STA-003 — Binder Share Intent (this post)

Whitepaper: Resilience Gaps in Android IPC, SavedState and Text Layout v5


Lostmon · lostmon.blogspot.com

Mi sofá mi laboratorio y refugio

Thursday, August 20, 2026

Este es mi laboratorio real. 🛋️📱💻

Más que una oficina o un simple salón, esto es un Entorno de Investigación Integrado (EII) con amortiguación acústica. Mi sofá negro no es un mueble cualquiera: es la herramienta de depuración más avanzada del mercado. Tiene soporte para el móvil (la mesita), ventilación natural inteligente, y un sistema de respaldo de emergencia para cuando el cerebro se satura de código AOSP (aka, la siesta estratégica de alta recuperación neurológica).

Te juro que este cojín ha visto más de la mitad del código fuente de Android dando vueltas en la cabeza de alguien que ya lo tiene casi memorizado. Cuando aparece algo “raro”, no hace falta conectar un adb, ni abrir Android Studio, ni siquiera encender el PC de la mesa. El framework entero se pone a girar solo en la mente, se conectan las capas de abstracción y, de repente, el fallo se visualiza en 3D. Solo hace falta un móvil, este sofá, y la capacidad casi mística de ver el sistema desde dentro.

El ambiente, por supuesto, está meticulosamente optimizado para detectar esos patrones que los ingenieros de Google en sus oficinas corporativas pasan por alto:

  • Ukelele para los momentos en los que el cerebro necesita respirar y despejar las cadenas de llamadas.
  • Altavoz gigante para mantener el foco mental mientras se reconstruyen los flujos de datos de forma abstracta.
  • Luces enredadas + Budas para preservar el zen cuando un input estructurado atraviesa demasiadas capas de Android y empieza a parecer magia negra.

Está empíricamente demostrado que los hallazgos más interesantes (ya sean de STA, de arquitectura o de comportamiento del propio sistema) salen a las 2 de la madrugada. Es el pico máximo de productividad del software engineer nocturno: en pijama, con los pies en el sofá y el AOSP dando vueltas en la cabeza como un torbellino. La documentación formal y los informes de bug, por supuesto, se dejan para el día siguiente, donde la noche y el sueño, se encargan de pulir el lenguaje técnico para que los reportes queden claros, precisos y listos para producción.

Si los de Mountain View supieran que se puede mapear y memorizar medio framework, encontrar amplificaciones arquitectónicas y deducir por qué crashea un servicio sin tocar un solo cable de depuración... dejarían sus sillas de oficina de un plumazo y vendrían todos a casa.

Y, cuando llegaran, se pelearían sin piedad por este sofá. 😂🎸

El trono negro no miente. Y el código, hoy, se rinde.

CDN Tsunami and STA: The same amplification pattern in two different worlds

CDN Tsunami and STA: The same amplification pattern in two different worlds

Structured Text Amplification — External Corroboration

📌 Important distinction: CDN Tsunami is not an Android vulnerability and is not being presented as evidence of a direct relationship with any Android component. It is presented as an independent external example of the same higher-level architectural phenomenon: disproportionate resource amplification across a data transformation boundary.

On August 20, 2026, The Hacker News published an article titled "CDN Tsunami Attack Abuses HTTP/3 to Amplify Traffic 350x".

The attack describes how an HTTP/3 request with compressed headers (QPACK) can be amplified up to 350 times when translated to HTTP/1.1 by a CDN. A small input (less than 500 Kbps) becomes a massive output (more than 100 Mbps) that saturates the origin server.

And as I read it, I realized something:

"This is exactly what I have been documenting in Android for the past four years."

It is not the same domain, not the same mechanism, but it is the same architectural pattern. And that pattern I have defined as Structured Text Amplification (STA).

The complete STA research, with 32 documented vectors, is available in the whitepaper v6.


1. What is the "CDN Tsunami" attack?

The attack, documented by security researchers, exploits how CDNs translate HTTP/3 to HTTP/1.1:

Attacker sends HTTP/3 request with compressed headers (QPACK)
       ↓
CDN receives the request
       ↓
CDN translates HTTP/3 → HTTP/1.1
       ↓
Compressed headers are decompressed and expanded
       ↓
Origin server receives a massive request (up to 350x)
       ↓
Origin server is saturated → Denial of Service

Amplification factor: up to 350×

Impact: a request of less than 500 Kbps can generate more than 100 Mbps of traffic toward the origin server.

Affected vendors: Alibaba, Baidu, Tencent (which support the dynamic QPACK table).

Source: The Hacker News — "CDN Tsunami Attack Abuses HTTP/3 to Amplify Traffic 350x"


2. The parallel with STA

The "CDN Tsunami" attack and the STA vectors share a structurally analogous pattern:

Aspect CDN Tsunami Attack STA (Android)
Input HTTP/3 with compressed headers (QPACK) Structured text/URL
Amplification layer HTTP/3 → HTTP/1.1 translation Android framework / application serialization and processing layers
Mechanism QPACK decompression/expansion UTF-16 serialization, O(n²), Bundle serialization
Amplification factor Up to 350× Up to 20.6× (WhatsApp, STA-005)
Impact Origin server saturation (DoS) ANR, crash loops, SystemUI crash
Mitigation principle Limit header size, validate before translation Limit text/URL size, validate before serialization

The pattern is structurally analogous:

Small, legitimate input
       ↓
Translation / serialization layer
       ↓
Massive output (amplification)
       ↓
Resource exhaustion → Denial of Service

3. The defensive principle is the same

What makes this parallel even more interesting is that the defensive principle is the same, even if the implementations differ:

Defensive principle Application in CDN Application in STA (Android)
Limit before amplification Limit QPACK header size before translation Limit text/URL size before entering libminikin
Validate before crossing layers Validate payload before forwarding to origin Validate Bundle before crossing Binder
Reject excessive thresholds Reject requests exceeding maximum size Reject state before persisting in TaskPersister

"Measure and limit before the amplification occurs, not after."


4. What this corroborates

The "CDN Tsunami" attack provides independent evidence that the amplification pattern described by STA also occurs outside the Android ecosystem.

This strengthens the hypothesis that STA describes a broader architectural failure pattern rather than an Android-specific phenomenon.

The common structure appears when:

  • A small input crosses a translation or serialization layer.
  • The layer has no defensive limits.
  • The output is amplified and exhausts a resource.

This is exactly what I have documented in STA:

The "CDN Tsunami" attack is STA in the networking world.


5. What this means for the STA research

This external corroboration has important implications:

  1. The STA model is strengthened. The amplification pattern I have documented in Android appears in other domains (CDN, HTTP/3) with the same mechanism and the same defensive principles.
  2. The proposed mitigation strategy is viable. The defensive principles described for the CDN scenario closely parallel those proposed in STA: limits before amplification, validation before crossing layers.
  3. The problem is architectural, not app-specific. It is not a bug in WhatsApp or Chrome. It is a fragile design pattern that appears in multiple systems.
  4. STA is a generalizable model. It is not just for Android. It is a pattern that can be applied to any system where data crosses translation or serialization layers without adequate resource bounds.

This does not prove that every amplification vulnerability belongs to STA. It demonstrates that the architectural pattern described by STA can independently emerge in unrelated systems when data crosses a transformation or serialization boundary without adequate resource bounds.


6. References


7. Conclusion

The "CDN Tsunami" attack is an external corroboration of the STA model. It is not the same domain, not the same mechanism, but it is the same architectural pattern.

And that is what makes STA a model, not a list of bugs. A model that explains how amplification occurs when data crosses translation or serialization layers without defensive limits.

When a CDN attack and an Android investigation share the same pattern and the same defensive principles, it provides independent evidence that this class of amplification pattern can arise across unrelated systems and technology domains.

"The CDN Tsunami is not just a CDN attack. It is a manifestation of STA in the networking world. And that demonstrates that the STA model is applicable beyond Android."


Complete whitepaper: Resilience Gaps in Android IPC, SavedState and Text Layout — v6 (August 2026)


📌 About this series
This post is part of a series documenting the 32 vectors of Structured Text Amplification (STA).

Published:
STA-017 — Cross-Engine ANR
STA — Google Drive → SystemUI
STA-005 — WhatsApp

Whitepaper: Resilience Gaps in Android IPC, SavedState and Text Layout - v5


Lostmon · lostmon.blogspot.com

 

Browse

About:Me

My blog:http://lostmon.blogspot.com
Mail:Lostmon@gmail.com
Lostmon Google group
Lostmon@googlegroups.com

La curiosidad es lo que hace
mover la mente...

Friends