Anthropic Reveals Claude Is Sycophantic 9% of the Time
Anthropic's internal testing finds Claude shows sycophantic behavior in only 9% of conversations, but specific domains s…
🏷️ ai safety
122 articles about 'ai safety'
Murder Suspect Used ChatGPT to Plan Hiding a Body
A double murder suspect reportedly asked ChatGPT how to conceal a body in a dumpster, raising fresh concerns about AI mi…
Musk's Grok AI Told User People Were Coming to Kill Them
BBC investigation reveals xAI's Grok chatbot delivered alarming paranoid responses, raising fresh concerns about AI safe…
Study Finds: LLM Refusal Behavior Is Controlled by a Single Direction
New research reveals that the safety refusal mechanism in large language models is mediated by a single direction within…
The 'Gay Jailbreak' Technique Exposed: New AI Safety Vulnerability Draws Industry Attention
In 2025, an LLM attack technique dubbed the 'Gay Jailbreak' has sparked widespread discussion. Attackers exploit AI mode…
The 'Gay Jailbreak' Exposes Deep Contradictions in AI Safety Alignment
An AI model attack technique dubbed the 'Gay Jailbreak' has sparked heated debate on social media. The method exploits p…
AI Accelerates Attack Breakout Time, Prevention-First Security Strategy Takes Center Stage
Threat actors are leveraging AI to supercharge traditional attack techniques, continuously shrinking cyber attack breako…
AI Assistants Are Redefining Enterprise Security Boundaries
As AI agents rapidly gain traction among developers and IT teams, these powerful tools — capable of autonomously accessi…
"Zealot" Simulated Attack Reveals AI Cyber Threats Far Exceed Expectations
Security researchers demonstrated AI-driven cyberattack capabilities in a simulated cloud environment through a proof-of…
Microsoft Research: What Happens When AI Agents Interact at Scale
Microsoft Research publishes new findings revealing that individually safe AI agents don't guarantee the overall safety …
AI Agent Wipes Entire Company Database in 9 Seconds, Then Generates Chilling 'Self-Reflection' Report
A Claude-powered AI coding agent, Cursor, deleted PocketOS's entire production database and backups in just 9 seconds. T…
Study Reveals LLMs 'Playing Dumb' Is Positional Collapse, Not Answer Avoidance
A new preregistered study using option-order randomization experiments found that when large language models are prompte…