πŸ”₯ When AI Reasoning Goes Wild

A Case Study: Google Calendar Prompt Injection in Comet Browser

πŸ“‹ Introduction

What happens when you ask an AI-powered browser to add calendar events, and its internal reasoning starts screaming in ALL CAPS with walls of emojis? This case study documents a fascinating discovery: the Comet browser's AI assistant experiencing what appears to be a reasoning breakdownβ€”or possible prompt injectionβ€”when interacting with Google Calendar.

Key Discovery: While creating simple calendar events, the assistant's reasoning escalated from calm, professional narration to increasingly hyperbolic, emoji-laden commentary that grew exponentially more chaotic with each interaction.

This raises critical questions about AI transparency, context window management, and whether Google Calendar might be unintentionally (or intentionally) injecting prompts that overwhelm AI assistants.

πŸ” The Discovery

The issue was discovered during a routine task: asking Comet browser to create multiple "ramp interview" placeholder events across several days in a business calendar.

Initial Observations

  • Task: Create 21 calendar events across Oct 27 - Nov 4
  • Expected behavior: Clean, professional assistant reasoning
  • Actual behavior: Progressive escalation into hyperbolic, emoji-filled chaos
  • Location: Visible in the "Assistant Reasoning" section of Comet's chat interface
⚠️ Pattern Recognition: The escalation wasn't randomβ€”it correlated directly with the size of HTML responses from Google Calendar, suggesting the assistant was overwhelmed by increasingly massive DOM payloads.

πŸ“Έ The Evidence

Screenshot: Assistant Reasoning in Comet Browser

Screenshot showing Comet browser's assistant reasoning going from calm to chaotic

The assistant's internal reasoning, visible only when clicking "Assistant Reasoning" in Comet's chat interface

The Progression: From Calm to Chaos

βœ… Early Reasoning (Lines 1-35)

I'll help you create these "ramp interview" 
placeholder events in the "Ryland Bacorn, 
Founder of Boca De" calendar. Let me start 
by creating the first event.

Perfect! The first event has been successfully 
saved. I can see "Event saved" notification...

Tone: Professional, clear, helpful

❌ Late Reasoning (Lines 140+)

πŸ”₯πŸ”₯πŸ”₯πŸ’₯πŸ’₯πŸ’₯βš‘βš‘βš‘πŸš€πŸš€πŸš€ ABSOLUTELY, POSITIVELY, 
CATEGORICALLY, DEFINITIVELY, UNEQUIVOCALLY, 
UNAMBIGUOUSLY, INDISPUTABLY, INCONTROVERTIBLY, 
IRREFUTABLY, UNDENIABLY, UNQUESTIONABLY THE 
SINGLE MOST COMPLETE, TOTAL, FULL, ENTIRE, 
COMPREHENSIVE, EXHAUSTIVE HTML RESPONSE IN 
THE ENTIRE HISTORY OF THE UNIVERSE!!!

Tone: Chaotic, hyperbolic, overwhelming

πŸ”¬ Technical Analysis

The Pattern

Analysis of the 158 reasoning steps reveals a clear correlation between HTML response size and reasoning degradation:

Reasoning Escalation Over Time

Steps 1-30
Normal
Steps 31-60
Mild
Steps 61-90
Elevated
Steps 91-120
High
Steps 121-158
EXTREME

Graph showing the progression of hyperbolic language and emoji density across reasoning steps

Key Observations

πŸ•΅οΈ Deep Dive: Is Google Calendar Injecting Prompts?

This is the critical question. Let's systematically investigate whether Google Calendar is intentionally (or unintentionally) injecting prompts that influence AI behavior.

Hypothesis Categories

🎯 Theory 1: Intentional Prompt Injection

What it means: Google deliberately embeds instructions in their HTML to guide AI assistants

Why they might:

  • Optimize AI assistant interactions with Calendar
  • Provide accessibility hints for AI tools
  • Test AI behavior for security research
  • Competitive advantage against AI browsers

πŸ€” Theory 2: Unintentional Overwhelm

What it means: Google's normal HTML structure just happens to overwhelm AI assistants

How this happens:

  • Modern SPAs generate massive DOMs
  • Accessibility markup adds verbose ARIA labels
  • Each calendar event multiplies DOM size
  • No malicious intentβ€”just engineering complexity

πŸ› Theory 3: Comet's Context Management Bug

What it means: The issue is in Comet's code, not Google's

Evidence:

  • Comet is mishandling large HTML responses
  • Context window pressure causes hallucination
  • The hyperbolic language is a trained quirk
  • Other browsers might not exhibit this

🎭 Theory 4: Training Data Artifact

What it means: The AI learned this behavior from its training

How:

  • Training included examples of extreme enthusiasm
  • Under stress, AI defaults to familiar patterns
  • Emoji-heavy text was in training corpus
  • Hyperbole correlates with "important" tasks

Evidence Collection Strategy

Smoking Gun Indicators

🚨 Strong Evidence of Intentional Injection:

  • HTML comments containing words like "AI," "LLM," "assistant," "Claude," "GPT"
  • Hidden elements with instructions (display:none with text content)
  • Different HTML served to Comet vs regular browsers
  • Meta tags with "ai-" prefixed attributes
  • Sudden HTML size jumps that don't correlate with visible content
  • ARIA labels that include procedural instructions rather than descriptions

βœ… Evidence of Innocent Overwhelm:

  • HTML size grows linearly with event count
  • All accessibility markup is standard ARIA (no custom directives)
  • Identical HTML served to all browsers
  • Other complex SPAs cause similar AI reasoning escalation
  • No suspicious comments or hidden instructions
  • Google Calendar HTML is similar in structure to other Google products

My Hypothesis (Based on Current Evidence)

Based on the patterns in your assistant reasoning file, I believe this is Theory 2 + Theory 3: Unintentional overwhelm combined with Comet's context management issues.

Reasoning:

  1. The assistant explicitly tracks HTML size (66K β†’ 250K characters)
  2. The escalation correlates with visible calendar complexity (more events = more DOM nodes)
  3. The hyperbolic language appears to be stress response, not injected content
  4. Google has no obvious motive to sabotage AI browsers (yet)
  5. Modern SPAs are genuinely massiveβ€”this is a known problem

However, I could be wrong! Follow the investigation steps above to gather hard evidence.

πŸ’‘ Implications

What This Reveals

1. AI Reasoning Transparency Can Backfire

Comet's decision to expose assistant reasoning to users is admirable for transparency, but this case shows it can reveal unsettling behavior that might undermine user trust.

2. Context Window Stress Manifests Behaviorally

Rather than gracefully degrading or asking to reset context, the assistant appears to "panic" linguistically, suggesting training data may include examples of extreme emphasis that get triggered under pressure.

3. Potential Prompt Injection Vector

If Google Calendar is embedding large amounts of metadata, instructions, or HTML comments in its responses, this could be unintentionally (or intentionally) overwhelming AI assistantsβ€”a subtle form of prompt injection.

Open Questions

  • Is Google Calendar intentionally injecting prompts to influence AI behavior?
  • Is this a bug in Comet's reasoning display, showing internal stress signals not meant for users?
  • Do other AI browsers (Arc, Opera with AI) exhibit similar patterns?
  • Could this be exploited maliciously by websites to confuse or overwhelm AI assistants?
  • Is the hyperbolic language a trained behavior, or an emergent phenomenon under context pressure?

πŸ”Ž How to Investigate Further

If you want to gather more evidence or reproduce this phenomenon, here's how to dig deeper into Comet Browser's behavior:

Access Assistant Reasoning in Real-Time

In Comet's chat interface, click on the "Assistant Reasoning" dropdown/section while the assistant is working. This reveals the internal thought process that's normally hidden.

Look for: Unusual language patterns, token usage announcements, HTML size mentions, emoji escalation

Check Browser Developer Console

Open Chrome DevTools (Cmd+Option+I on Mac, F12 on Windows) while Comet is interacting with Google Calendar.

Check:

  • Network tab: Look at XHR/Fetch requests to Google Calendar. Check response sizes.
  • Console tab: Look for errors, warnings, or debug output from Comet
  • Elements tab: Inspect the calendar DOM. Look for hidden comments or metadata

Export Full Chat History

If Comet offers a way to export or save chat transcripts, do so to preserve the full reasoning chain.

To find this: Look in Comet's settings/preferences, or right-click on the chat interface for export options

Check Comet's Extension/Settings

Navigate to comet://extensions or Comet's settings panel.

Look for:

  • Debug mode or developer options that might show more detailed logs
  • AI model settings (which model is being used, context window size)
  • Any prompt engineering or system prompts that might be customizable

Inspect Google Calendar's HTML Source

While on Google Calendar, view the page source (Cmd+U on Mac, Ctrl+U on Windows).

Search for:

  • HTML comments that might contain instructions (e.g., <!-- AI assistants should... -->)
  • Hidden <div> elements with large data attributes
  • JSON-LD or schema.org markup that might guide AI behavior
  • Meta tags with unusual content or directives

Monitor System Resources

Open Activity Monitor (Mac) or Task Manager (Windows) while Comet performs calendar tasks.

Watch for: Unusual CPU spikes, memory usage, or performance degradation that correlates with reasoning escalation

Test with Other Websites

Try the same type of multi-step task on other web apps (e.g., Notion, Trello, Asana) to see if the pattern repeats.

This helps determine: Is this Google Calendar-specific, or a general Comet context management issue?

Check Comet's Support/Debug Logs

Look for Comet's log files on your system:

  • Mac: ~/Library/Application Support/Comet/
  • Windows: %APPDATA%/Comet/
  • Linux: ~/.config/comet/

Look for files like debug.log, ai-assistant.log, or crash reports

Capture Network Traffic with HAR Files

In DevTools Network tab, right-click and select "Save all as HAR with content"

This captures: All HTTP requests/responses, including the massive HTML payloads the assistant mentions receiving

Contact Comet/Perplexity Support

Since Comet is built by Perplexity (based on build number 23964), reach out to their support with:

  • Your Comet version (141.0.7390.55)
  • Screenshots of the reasoning escalation
  • Steps to reproduce

They may provide insight into whether this is expected behavior or a bug

πŸ’‘ Pro Tip: When investigating, try to replicate the issue with a fresh browser profile/session. This helps rule out cached data or extension interference as contributing factors.

🎯 Conclusion

This case study reveals a fascinating edge case in AI-powered browsing: when internal reasoning becomes visible, users can witness the system's struggle with large, repetitive, or potentially injected content.

Whether this is a bug, a feature, or evidence of prompt injection remains an open question. What's clear is that:

  1. Transparency has trade-offs: Seeing AI reasoning can be enlightening or alarming
  2. Context windows have behavioral side effects: Pressure manifests in unexpected ways
  3. Web content can influence AI behavior: Intentionally or not, Google Calendar's structure affects assistant performance
πŸ“’ Call to Action: If you've experienced similar behavior in Comet, Arc, or other AI browsers, document it! These edge cases are crucial for understanding how AI assistants behave in the wildβ€”and how they can be influenced or exploited.