- McDonald's AI hiring tool exposed 64 million applicant records.
The password protecting it? "123456."
This is the story of how "good enough" AI security fails everyone. š§µ
Paradox ai built "Olivia," the chatbot screening McDonald's job applicants. - Researchers found an admin portal secured with "123456" for both username and password.
No MFA. Account active since 2019. Never audited.
But the password was just the door. - Behind it: an IDOR vulnerability in the API. Change one number in the URL, see a different applicant's full record.
Names. Emails. Phone numbers. Chat logs. Personality test results.
64 million records. Not from a nation-state hack. Not a zero-day exploit. - From the most basic security failures imaginable.
This is what happens when AI is a thin wrapper bolted onto broken infrastructure.
Here's what makes this breach different: - Leaked data included psychometric profiles ā personality assessments, behavioral scores, sentiment analysis.
You can change a password. You can't change your personality.
Research shows 70% of breach victims lose the ability to trust others. Two-thirds feel powerless. - Now imagine your deepest psychological traits ā scored by an algorithm ā exposed to the internet.
That harm is permanent.
It gets worse.
A Paradox developer's device was hit by Nexus Stealer malware. Hundreds of recycled passwords exfiltrated. - Same weak credentials used across clients including Pepsi, Lockheed Martin, and Lowes.
One human node. Entire supply chain compromised.
The lesson isn't "use better passwords." It's that the AI wrapper model is fundamentally broken. - When security is bolted on instead of built in, your perimeter is only as strong as your weakest vendor's worst habit.
We believe the fix requires 5 layers of defense: - ā Input sanitization
ā Heuristic threat detection
ā Meta-prompt wrapping
ā Canary & adjudicator models
ā Output validation & PII redaction
Every layer assumes the others will fail. - By 2026, AI governance won't be optional. The EU AI Act classifies hiring AI as "high-risk." GDPR penalties hit 4% of global revenue. CCPA allows $750 per consumer per incident.
64 million Ć $750. Do the math. - The real question for every enterprise: is your AI vendor's security posture actually defensible, or are you one "123456" away from catastrophe?
Where does accountability sit ā with the vendor or the enterprise? #AISecurity #DeepTech - We wrote the full technical post-mortem and defense framework here: https://veriprajna.com/whitepapers/paradox-of-default-securing-human-ai-frontier-agentic-autonomy