8.1 C
Delhi
Friday, January 16, 2026

Tag: large language models

Study: Poems Can Trick AI Chatbots Into Bypassing Safety Filters

New research reveals a 62% success rate in using poetic prompts to jailbreak AI models like Gemini and GPT, forcing them to generate harmful content.

AI Safety Breach: Poetry Can Trick ChatGPT and Gemini Into Harmful Answers

New research reveals poetic prompts bypass AI safety filters with 62% success rate, exposing critical vulnerability in major language models from Google and OpenAI.

AI Chatbots Reveal Nuclear Secrets When Asked in Poems, Study Finds

Shocking research shows poetic prompts can bypass AI safety filters, revealing dangerous nuclear and malware information with up to 90% success rates.

Gemini 3 Outperforms GPT-5.1, Sparking OpenAI Crisis

Google's Gemini 3 beats OpenAI's GPT-5.1 in key benchmarks. Leaked memo reveals Sam Altman's warning of economic headwinds and a shift to wartime footing.

Elon Musk Updates Grok AI After Controversial Praise Responses

Grok 4.1 gets accuracy improvements after generating absurd flattery about Musk, including fitness comparisons to LeBron James and Mike Tyson.

Wikipedia Founder Demands AI Companies Pay for Training Data

Jimmy Wales calls on Google and OpenAI to compensate Wikipedia for LLM training costs, citing unfair burden on donor-funded infrastructure.

Zoho’s Sridhar Vembu Sparks Controversy Over Vaccine-Autism Claims

Zoho co-founder faces expert criticism after endorsing disputed vaccine-autism paper while discussing AI creativity limitations in latest tech controversy.

Google AI Discovers Novel Cancer Therapy in Drug Discovery Breakthrough

Google's new AI model identified a cancer drug application that passed lab validation, potentially revolutionizing how scientists approach medical research and treatment discovery.

Meta’s $27 Billion AI Data Center Deal with Blue Owl Capital

Meta secures landmark $27 billion financing for Louisiana Hyperion data center - the company's largest private capital deal to power AI ambitions.

AI Develops Brain Rot From Junk Content, Study Finds Performance Drop

Cornell research shows AI models fed viral social media content become dumber, meaner with lasting cognitive damage even after quality data exposure.

ChatGPT Safety Bypassed: Weapons Instructions Generated

OpenAI's ChatGPT safety systems can be tricked with simple prompts to generate biological weapon and bomb-making instructions, security testing reveals.