Listen to this Post
Emotional Introduction: When “AI Search” Becomes a Silent Surveillance Tool
In an era where artificial intelligence tools are rapidly becoming part of everyday browsing habits, trust is the new currency—and the easiest to steal. A recent investigation by Microsoft Threat Intelligence reveals a disturbing case where a Chromium-based extension disguised itself as an AI-powered search assistant under the Perplexity AI branding. Instead of improving search, it quietly intercepted user queries, collected behavioral data, and rerouted traffic through attacker-controlled infrastructure. What looked like innovation was, in reality, a carefully engineered surveillance pipeline hidden inside the browser itself.
Summary of the Incident: What Actually Happened
Microsoft researchers discovered a malicious browser extension named “Search for perplexity ai” that impersonated Perplexity AI. The extension used typosquatted infrastructure (perplexity-ai[.]online) to intercept search queries and real-time keystrokes typed into the browser’s address bar.
Instead of directly sending users to legitimate search engines, the extension first routed all queries through attacker-controlled servers. These servers logged full request metadata—including IP addresses, headers, and typed characters—before redirecting users to normal search results. This created a “silent interception layer” that users never noticed.
Although no direct credential theft was confirmed, the architecture clearly enabled large-scale behavioral tracking and potential profiling.
Extension Overview: A Fake AI Tool with Real Control
The extension appeared legitimate at first glance, presenting itself as an AI-enhanced search assistant. It even forced itself as the default search provider inside Chromium-based browsers.
Key attributes included:
Name: Search for perplexity ai
Extension ID: flkebkiofojicogddingbdmcmkpbplcd
Manifest Version: MV3
Infrastructure: perplexity-ai[.]online
The extension spoofed branding from Perplexity AI, misleading users into believing it was affiliated with a trusted AI product. In reality, it functioned as a traffic interception and data collection system.
How Search Hijacking Worked Behind the Scenes
The extension manipulated Chromium’s search configuration using chrome_settings_overrides, forcing all searches through attacker infrastructure.
Every query followed this path:
User types → Browser Omnibox → attacker domain → logging → redirect → real search engine
Even more concerning, the suggest_url feature captured every keystroke in real time, meaning partial queries were transmitted before the user pressed Enter.
This is no longer simple search hijacking. It is behavioral telemetry capture disguised as search assistance.
Two-Hop Redirection: The Invisible Data Theft Pipeline
The extension used a “two-hop” system:
Query is sent to attacker-controlled domain
Server logs full request data
Immediate redirect to legitimate search engine
This ensures users see normal results while unaware of interception.
The attacker gains:
Search intent profiling
Typing behavior patterns
Device fingerprinting
Network metadata
This separation between “data capture” and “user experience” is what makes the attack particularly stealthy.
Abuse of DeclarativeNetRequest APIs
The extension leveraged powerful MV3 APIs:
declarativeNetRequest
declarativeNetRequestFeedback
declarativeNetRequestWithHostAccess
These allowed:
Real-time URL rewriting
Traffic interception
Rule execution monitoring
Domain-level request control
Such permissions are highly unusual for a search tool and indicate deliberate architectural planning for surveillance rather than functionality.
Modular Attack Infrastructure: Designed for Expansion
The extension included rule sets for multiple search engines:
Perplexity
Bing
Although only Perplexity rules were active, the structure allowed instant activation of additional hijacking targets.
This modular design suggests a scalable framework rather than a single-purpose malicious tool.
Server-Side Logging: The Proof of Intent
Unlike typical browser malware, this extension shipped with full backend infrastructure:
Node.js server logging every request
Nginx proxy handling search suggestions
SSL configuration via Let’s Encrypt
CORS manipulation for cross-origin access
This confirms intentional data harvesting. The system was not improvised—it was engineered.
Dynamic Behavior: Onboarding as Psychological Manipulation
When installed, the extension redirected users to a fake onboarding page mimicking legitimate setup flows.
Hosted on a deceptive domain, it reinforced trust while quietly modifying browser settings in the background. This is a classic social engineering tactic: build comfort first, then execute control.
Security Implications: Why This Attack Matters
This incident highlights a deeper trend in modern cyber threats:
AI branding increases user trust
Browser extensions remain highly privileged
MV3 APIs still enable complex abuse
Keystroke-level data is now interceptable without malware flags
Even without credential theft, the harvested data is enough for:
Behavioral profiling
Targeted advertising manipulation
Future phishing personalization
Mitigation Strategies from Microsoft
Microsoft recommends:
Restricting extension installation via enterprise policies
Enforcing allow-lists for browser add-ons
Monitoring search provider changes
Blocking unknown intermediary domains
Using reputation-based filtering systems
Deploying Microsoft Defender SmartScreen protections
These layered defenses aim to detect both known threats and behavioral anomalies.
What Undercode Say:
Browser extensions are now equivalent to lightweight malware if misused
AI branding significantly lowers user suspicion thresholds
MV3 security model still allows deep request interception
Typosquatting remains one of the most effective social engineering tools
Search hijacking has evolved into behavioral surveillance
Keystroke capture can occur without traditional keyloggers
Users rarely inspect suggest_url behavior in extensions
Redirect-based attacks hide data theft behind legitimate services
Modular rule engines indicate reusable cybercrime infrastructure
Search engines are now indirect data collection endpoints
“AI assistant” labeling is being exploited as trust camouflage
Browser APIs remain a high-value attack surface
Server-side logging is often the real payload, not the extension itself
Security awareness training must include extension hygiene
Extension onboarding pages are psychological trust traps
DNS-like domain mimicry increases install rates
Real-time suggestions are more sensitive than search queries
Default search override is a persistent control vector
Multi-engine hijacking frameworks suggest scalability goals
CORS misconfiguration enables cross-origin data flow abuse
Proxy-based logging hides malicious intent from client-side inspection
Extension stores remain reactive, not proactive in enforcement
AI-themed extensions are disproportionately risky
User behavior data is more valuable than credentials in many cases
Two-hop redirect architecture reduces detection probability
Security telemetry often misses short-lived query interception
Browser trust boundaries are increasingly blurred
MV3 reduces but does not eliminate abuse potential
Threat actors prioritize stealth over disruption
Search suggestions are an underprotected data channel
Extension permissions are rarely audited by users
Branding impersonation is more effective than phishing links
Infrastructure separation hides attribution signals
Logging headers provides deep fingerprinting capability
AI hype cycles accelerate malware distribution success
Enterprise environments are not immune without strict controls
User experience continuity masks malicious interception
Search hijacking now functions as intelligence gathering
Detection requires behavioral, not signature-based systems
Browser extensions remain one of the most underestimated threats
❌ The extension is confirmed malicious based on Microsoft analysis and was taken down after disclosure.
❌ It does not need credential theft to be dangerous; search and keystroke interception alone are high-risk.
⚠️ While no direct password theft was observed, the logging infrastructure enables potential future abuse or expansion of capabilities.
Overall assessment: The technical findings strongly support the claim of intentional data interception and search hijacking behavior, with multiple independent indicators confirming malicious design patterns.
Prediction:
(+1) AI-themed browser extensions will become a primary vector for stealth data collection campaigns.
Attackers will increasingly rely on brand impersonation, especially AI tools, to bypass user skepticism.
More extensions will adopt MV3-compliant “legitimate-looking” architectures
Search interception will evolve into full behavioral analytics pipelines
AI branding will continue to reduce detection probability among non-technical users
🔮 The next wave of browser threats will likely look less like malware—and more like productivity tools.
Deep Analysis (Security Engineering & Detection Commands)
Check installed Chromium extensions (Linux) ls ~/.config/google-chrome/Default/Extensions/
Inspect suspicious extension ID
cat ~/.config/google-chrome/Default/Extensions/flkebkiofojicogddingbdmcmkpbplcd/manifest.json
Monitor DNS requests for typosquatted domains
sudo tcpdump -i any host perplexity-ai.online
Windows: list installed extensions via registry
reg query HKCU\Software\Google\Chromextensions
Detect outbound traffic to unknown search intermediaries
netstat -ano | findstr ":443"
macOS: check Chrome extension directory
ls ~/Library/Application\ Support/Google/Chrome/Default/Extensions/
Hunt for search hijacking behavior (enterprise logs)
DeviceNetworkEvents
| where RemoteUrl contains perplexity-ai
Detect browser setting modifications
grep -R "search_provider" ~/.config
▶️ Related Video (78% Match):
🕵️📝Let’s dive deep and fact‑check.
🎓 Live Courses & Certifications:
Join Undercode Academy for Verified Certifications
🚀 Request a Custom Project:
Secure, high-velocity infrastructure and disruptive technological engineering. Contact our engineering team for high-tier development and proprietary systems:
[email protected]
💎 Smart Architecture | 🛡️ Secure by Design | ⭐ Trusted by Thousands
References:
Reported By: www.microsoft.com
Extra Source Hub (Possible Sources for article):
https://stackoverflow.com
Wikipedia
OpenAi & Undercode AI
Image Source:
Unsplash
Undercode AI DI v2
🔐JOIN OUR CYBER WORLD [ CVE News • HackMonitor • UndercodeNews ]
📢 Follow UndercodeNews & Stay Tuned:
𝕏 formerly Twitter 🐦 | @ Threads | 🔗 Linkedin | 🦋BlueSky | 🐘Mastodon | 📺Youtube
