3 Guys Hacked OpenAI Using a Rival Anthropic Model. Here's What It Shows About Frontier Labs' Vulnerabilities.

TL;DR

Cybersecurity researchers from Hacktron hacked OpenAI's community forum using Anthropic's AI models.
Advanced AI models act as 'force multipliers' for hackers, simplifying and accelerating cyberattacks.
The researchers gained access to ChatGPT and Codex accounts of some OpenAI users.
This access could have allowed them to view user conversations and potentially access connected applications like email and Slack.
The hack highlights the swift pace of AI development, where capabilities can change significantly overnight.
The researchers emphasized that they were acting as 'good guys' to identify security gaps.
Newer AI models could potentially allow such hacks to be completed in less than a day.
The Hacktron team has previously targeted other major tech companies like Apple, Google, and Facebook.
The exploit involved finding a flaw in a third-party service (Discourse) used by OpenAI, not directly in OpenAI's code.
OpenAI has acknowledged the breach, stating they have narrowed permissions and revoked affected tokens.
The incident raises concerns about national security implications of powerful AI models and the race to develop them.
Experts suggest AI labs may be prioritizing speed over robust security measures.

Continue reading the original article