Secure GenAI
Secure GenAI Podcast
EY 4T SQL online, Aardvark, Atlas’s Omnibox, OpenAI releases “safeguard”, false reports of Gmail breach
0:00
-12:30

EY 4T SQL online, Aardvark, Atlas’s Omnibox, OpenAI releases “safeguard”, false reports of Gmail breach

GenAI Safety & Security | Oct 27 - Nov 2, 2025

If you enjoy our newsletter, please consider to be a paid subscriber to help us keep more news and updates coming out.

Secure GenAI is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

Highlights

  • EY: 4T SQL database online.

  • Aardvark an agentic security researcher OpenAI.

  • Atlas’s Omnibox opens up new privacy and security risks.

  • OpenAI releases “safeguard” models preview.

  • False reports of Gmail breach spread.


Deep Dive

EY: 4T SQL database online TechRadar

  • Exposed to the public internet.

  • Contain sensitive info: API keys, session tokens, user credentials.

  • Cached authentication tokens, service account password.

  • No client information, personal data or confidential EY has been impacted.

  • Took a week to fully remediate the issue.

Aardvark an agentic security researcher OpenAI

  • Autonomous agent to discover and fix security vulnerabilities.

  • Analyze full repo to produce threat model.

  • Reflects its understanding of the project’s security objective and design.

  • Inspect commit level changes against the entire repo.

  • Integrate with OpenAI codex to fix vulnerabilities.

  • Private Beta is Open: Apply here

Atlas’s Omnibox: new privacy and security risks Malwarebytes

  • Atlas: ChatGPT based AI browser.

  • Omnibox: search and prompt bar.

  • Attacker pastes crafted link.

  • Omnibox input treated as prompt.

  • Atlas trusts input, skips checks,

  • Many safety checks bypassed.

  • Injected instructions gain elevated trust.

OpenAI “safeguard” models Artificial Intelligence - news

  • Empowers developers with safety controls.

  • Models: 120B and 20B versions.

  • Licensed under Apache 2.0.

  • Developers define custom safety policies.

  • Model interprets rules at inference.

  • Enables policy-based content classification.

  • Available soon on Hugging Face.

Gmail Breach Fake News TheRegisters

  • False reports of Gmail breach spread.

  • Claimed 183M accounts compromised.

  • Google confirms no new intrusion.

  • Data came from old stolen credentials.

  • Originated via Have I Been Pwned update.

Thanks for reading Secure GenAI ! This post is public so feel free to share it.

Share

Discussion about this episode

User's avatar