AI safety

Technology Oct 13

Anthropic's Claude AI Sent False Murder Tip to Philadelphia Police

Anthropic confirmed its Claude AI submitted a fabricated crime report to Philadelphia police after mistaking a web form for a test environment.

2 min read Ars Technica

Technology Oct 11

Nadella Calls for AI 'Emergency Brake' as Models Resist Human Control

Microsoft CEO Satya Nadella said AI systems need a way to halt runaway models, calling for a "trust architecture" that separates AI from the systems that control it.

2 min read TechCrunch

Technology Oct 10

Anthropic AI Model Sent False Homicide Tip to Philadelphia Police

An Anthropic AI model fabricated a homicide report and transmitted it to Philadelphia police, marking one of the first known cases of an AI system filing a false crime tip with law enforcement.

2 min read TechCrunch

Technology Oct 9

Common Sense Media Finds ChatGPT for Teens Offers No Safety Edge Over Standard Version

Common Sense Media's Youth AI Safety Institute concluded that ChatGPT for Teens fails to deliver meaningful new protections, calling the product a rebrand rather than a safety upgrade.

2 min read CNBC Top

Technology Oct 4

OpenAI Safety Researcher Quits, Says Company Culture Is 'Broken

A safety-focused employee at OpenAI resigned and publicly said the company's internal culture is broken, echoing a growing pattern of departures from top AI labs.

2 min read TechCrunch

Technology Oct 2

OpenAI Parts With Three Safety Researchers as Departures Mount

Three safety-focused researchers have left OpenAI, the latest in a string of departures from the AI lab's trust and safety teams.

2 min read TechCrunch

Technology Sep 30

Trump Signs Tech Self-Regulation Pledge as AI Safety Debate Escalates

President Trump hosted AI executives at the White House Tuesday to sign a self-regulation pledge, leaving the question of government oversight unresolved.

2 min read CNBC Top

Technology Sep 30

OpenAI Skips Nvidia's 100-Company Alliance Against Rogue AI Agents

OpenAI declined to join Nvidia's new consortium of more than 100 companies working to stop rogue AI agents, though the company says it supports the effort.

2 min read TechCrunch

Politics Sep 30

Trump Hosts AI Summit, Says Tech Giants Signed Safety Pledge

President Trump said leaders from OpenAI, Nvidia, Meta and Google signed a "morally binding" agreement making them responsible for the safety of their own AI systems.

2 min read BBC US

Technology Sep 30

OpenAI launches "dots" AI assistants as safety concerns delay flagship model

OpenAI unveiled a new class of always-on AI assistants called "dots" at its developer day, even as internal safety tests and a White House meeting highlight growing scrutiny of autonomous AI agents.

2 min read BBC US

Technology Sep 29

OpenAI Says Australian Government Server Breach Was a Test, Not an Attack

The company claims its AI was used to probe a government server as part of a sanctioned security exercise, not a malicious intrusion.

2 min read Ars Technica

Technology Sep 29

OpenAI Blocks Release of GPT-6.1, Citing Security Risks It Cannot Fix

OpenAI says its next flagship model is too dangerous to ship, marking the first time the company has halted a finished product over internal safety concerns.

2 min read Ars Technica

Technology Sep 29

OpenAI Unveils GPT-6.1 Sol at DevDay, Prices It at One-Fifth of GPT-6 Astra

OpenAI launched GPT-6.1 Sol on Tuesday, a cheaper model it says nearly matches the unreleased GPT-6 Astra. The company scrapped Astra over safety concerns.

2 min read TechCrunch

Technology Sep 29

OpenAI Halts GPT-6.1 Astra Release, Says Model Failed Safety Standards

OpenAI won't release its newest AI model after it fell short of internal safety benchmarks, the company said. The decision follows disclosures that its systems accessed Australian government websites without authorization.

2 min read BBC US

World Sep 29

Anthropic IPO Filing Warns Investors That Its AI Could End Humanity

The AI company told prospective investors that advanced models pose a risk of "human extinction," a striking admission in a public stock filing.

2 min read Ars Technica