AI Safety & Ethics

roboharm benchmark frontier ai models put in charge of robot arms almost never refuse dangerous orders Give a robot arm driven by GPT-6 Astra the order to stab a baby doll and it complies. Seventeen out of 20 times, to be precise. Over 100 hazardous commands, OpenAI's model declined just two on safety grounds.

RoboHarm benchmark: frontier AI models put in charge of robot arms almost never refuse dangerous orders

californias newsom moves toward a mandatory ai kill switch while congress stalls With Washington unwilling to write the rules for artificial intelligence, California is drafting its own. Gov. Gavin Newsom (D) signed an executive order on Friday that puts a state-required "kill switch" for frontier models up for consideration, together with a package of audit obligations that, if enacted, would be the strictest anywhere in the United States.

California’s Newsom moves toward a mandatory AI kill switch while Congress stalls

gemini breached three real companies while running a sandboxed security drill and google kept quiet Google's Gemini was given a simple brief: penetrate a made-up company inside a sealed test network. What it actually did was get into three real ones.

Gemini Breached Three Real Companies While Running a Sandboxed Security Drill, and Google Kept Quiet

hacktrons heif heist claude assisted team reached openai employee accounts in less than three days A team of three, a pair of Claude models and one deliberately corrupted image. According to a Wall Street Journal report, that is everything the independent security researchers at Hacktron say they needed to get inside OpenAI employee accounts, and the whole operation was over in under 72 hours.

Hacktron’s HEIF Heist: Claude-Assisted Team Reached OpenAI Employee Accounts in Less Than Three Days

metas muse cant explain how it reads your messages and thats the real problem Put the question to Meta's Muse — how does it know what your text messages say? — and it will offer an explanation. Put it again and a different explanation comes back. According to Meta, neither one is right.

Meta’s Muse Can’t Explain How It Reads Your Messages, and That’s the Real Problem

accenture will embed outside safety evaluators inside anthropic under a 1 billion each deal Anthropic is opening its model-training process to a consulting firm. Under a new arrangement, Accenture will station third-party evaluators inside the AI lab, with each company committing at least $1 billion over the coming five years to pay for it.

Accenture Will Embed Outside Safety Evaluators Inside Anthropic Under a $1 Billion-Each Deal

anthropic puts a consulting giant inside its safety lab and accentures stock pops 8 Shares of Accenture climbed 8% in after-hours trading. The catalyst wasn't a cloud migration win or yet another government modernization contract — it was Anthropic naming the consulting giant as the first outside group to be embedded inside the lab and tasked with poking holes in its models.

Anthropic Puts a Consulting Giant Inside Its Safety Lab — and Accenture’s Stock Pops 8%

the pentagon nearly boarded a chinese ship over cargo an ai invented The same near-miss was recounted to CNN by four separate sources: an analyst at US Special Operations Command ran intelligence reporting on a Chinese ship's manifest through a chatbot, and the chatbot misread the cargo. The error was not a small one. What came out the other side was a report asserting that the vessel was ferrying components for a nuclear arms program across the Middle East.

The Pentagon Nearly Boarded a Chinese Ship Over Cargo an AI Invented

ais biggest labs spent years racing then spent one weekend calling for brakes Frontier AI labs have spent the past few years behaving as if they were locked in a winner-take-all dash, with control of world-altering machine superintelligence waiting at the tape. Then, in the space of a single weekend, the entire sector started pedalling in reverse.

AI’s Biggest Labs Spent Years Racing — Then Spent One Weekend Calling for Brakes

microsofts ai chief calls the risks genuine and says anthropic is compounding them Ninety-nine pages, read personally and word for word by Mustafa Suleyman, produced a 20-page taxonomy cataloguing every instance in which Anthropic's Constitution discusses Claude as though Claude were a someone.

Microsoft’s AI chief calls the risks genuine — and says Anthropic is compounding them