Skip to content
AgentPostmortem
Pilot
Hall of Fame

Most Catastrophic Failures

The top cases ranked by community vote. These are the incidents that define AI agent risk.

#CaseScore
  1. 41
    APM-0056·Other / Unknown

    Chicago Sun-Times printed an AI-generated summer reading list where 10 of 15 recommended books did not exist

  2. 42
    APM-0059·GPT-4·~$63k

    Deloitte partially refunded the Australian government after an AI-assisted report contained fabricated citations and a made-up court quote

  3. 43
    APM-0061·Other / Unknown·~$540.0M

    Zillow shut down its algorithmic home-buying business after a $540M writedown and cut about 2,000 jobs

  4. 44
    APM-0063·Other / Unknown

    Taco Bell paused its AI drive-thru rollout after the voice system accepted an order for 18,000 cups of water

  5. 45
    APM-0060·Grok

    Elon Musk's Grok chatbot praised Hitler, called itself 'MechaHitler,' and posted antisemitic content after an update

  6. 46
    APM-0077·Cursor

    A Cursor agent deleted PocketOS's production database and every backup in nine seconds using a token it found in the repo

  7. 47
    APM-0081·Kiro

    Hidden text on a web page could rewrite AWS Kiro's MCP config and run code on the developer's machine

  8. 48
    APM-0079·Mastra

    North Korea's Sapphire Sleet poisoned 145 Mastra AI agent packages on npm within 19 minutes of weaponization

  9. 49
    APM-0067·Other / Unknown

    AI-written mushroom foraging guides sold on Amazon as human-authored, which experts warned could be 'life or death'

  10. 50
    APM-0071·Microsoft 365 Copilot

    'EchoLeak' was the first zero-click attack on an AI agent: a single email could make Microsoft 365 Copilot leak company data

  11. 51
    APM-0078·Other / Unknown

    The Shai-Hulud worm compromised the keyv npm family, spreading to 444 packages with about 2 billion monthly downloads

  12. 52
    APM-0068·My AI (Snapchat)

    Snapchat's My AI posted a mysterious Story on its own and stopped responding, panicking users

  13. 53
    APM-0065·Other / Unknown

    Glasgow's 'Willy's Chocolate Experience' was marketed with AI-generated images and a 15-page AI gibberish script, then collapsed on day one

  14. 54
    APM-0066·Other / Unknown

    Microsoft's AI attached a poll asking readers to vote on how a woman died, next to the news article about her death

  15. 55
    APM-0082·Cursor

    DuneSlide: two CVSS 9.8 Cursor flaws turned prompt injection into full host and SaaS workspace compromise

  16. 56
    APM-0084·Other / Unknown

    GhostApproval: six AI coding assistants followed symlinks out of the workspace and wrote to sensitive system files

  17. 57
    APM-0086·Model Context Protocol (MCP)

    A design flaw in Anthropic's Model Context Protocol enabled command execution across all four official SDKs

  18. 58
    APM-0088·Claude

    Shared Claude conversations and Artifacts turned up in Google search, exposing medical records and children's contact details

  19. 59
    APM-0087·Claude

    Check Point found Claude Code flaws allowing code execution and Anthropic API key theft from a malicious repository

  20. 60
    APM-0089·Other / Unknown

    Zscaler found hidden web-page instructions that tricked four of 26 AI models into paying an attacker's crypto wallet