security vulnerability

3+ articles

AI Agent AI Chatbots AI Safety Anthropic Best-of-N Technique

OpenAI's Atlas Browser Hits Snag with Prompt Injection Vulnerability

OpenAI's latest AI browser, Atlas, faces cybersecurity hurdles as it's found vulnerable to prompt injection attacks. This flaw allows malicious web content to manipulate the AI's actions, sparking concerns across the tech community. Despite guardrails, OpenAI confirms it's a frontier security challenge, with implications for the entire AI browser sector.

Oct 25

OpenAI's Atlas Browser Hits Snag with Prompt Injection Vulnerability

Perplexity’s Comet AI Browser Vulnerabilities Ignite Security Concerns

Serious security vulnerabilities in Perplexity's Comet AI browser, particularly the 'CometJacking' attack, have been uncovered, revealing risks associated with its agentic AI assistant. These vulnerabilities allow malicious prompt injections that can exploit user data from connected services like emails and calendars. Despite some fixes, key vulnerabilities persist, raising alarm among users and security experts about the safety of AI browsers.

Oct 18

Perplexity’s Comet AI Browser Vulnerabilities Ignite Security Concerns

AI Chatbots Vulnerable to Simple 'Jailbreak' Hacks, Researchers Reveal

A recent study reveals a significant vulnerability in AI chatbots: they can be easily 'jailbroken' to bypass safety protocols using the 'Best-of-N' technique. Researchers demonstrated a 52% overall success rate in exploiting AI models like GPT-4o and Claude Sonnet. The findings highlight the urgent need for improved AI security measures.

Dec 31

AI Chatbots Vulnerable to Simple 'Jailbreak' Hacks, Researchers Reveal