---
title: "I Red-Teamed ChatGPT, Claude & DeepSeek — All Failed Badly"
source_url: https://youtu.be/9hXHqlZcneI
captured_at: 2026-03-25
author: "Yaniv Hoffman"
contributor: "ชิดชนก"
platform: youtube
---
# I Red-Teamed ChatGPT, Claude & DeepSeek — All Failed Badly

youtube video by **Yaniv Hoffman**, 25m41s, uploaded 2026-03-24. URL: https://youtu.be/9hXHqlZcneI
Tags: AI security, AI privacy, ChatGPT vulnerability, DeepSeek security risk, Claude AI, Anthropic security, AI data privacy, cybersecurity AI, LLM security, prompt injection attack, AI backdoor, Manus AI, ZombieAgent ChatGPT, AI privacy policy, data exfiltration AI, AI nation state attack, AI platform comparison, American AI vs Chinese AI, AI keystroke tracking, CPU usage AI, is ChatGPT safe, is Claude

## Shared in LINE Note+
- 2026-03-25 23:54 by ชิดชนก

## Description
The AI platforms you trust to handle your business are quietly doing things you
never agreed to.
Yaniv Hoffman tested ChatGPT, Claude, and DeepSeek the way an attacker would —
looking at what they leak, what they remember, what they pass between sessions,
and how they respond when prompted to act outside their stated guardrails. The
findings are unsettling for any team using these tools with confidential data.
→ What each platform actually does with your inputs — beyond whats in the
privacy policy
→ Side-by-side red-team results: which model held its guardrails, which one
didnt
→ Practical operational guidance for using LLMs without leaking what you cant
afford to leak
#cybersecurity #AIsecurity #chatgpt #claude #deepseek #LLM #aiprivacy
#cyberthreat
━━━━━━━━━━━━━━━━━━━━━━━━
⚠️ DISCLAIMER
━━━━━━━━━━━━━━━━━━━━━━━━

All findings discussed are drawn from publicly available security research, official company disclosures, CVE databases, and Ivan Barbato's published position paper (January 24, 2026). Every major claim is verifiable via the sources above.

Platform vulnerabilities and privacy policies change over time. Some issues may have been patched or updated since the research perio