
Microsoft CEO Satya Nadella Calls for Emergency Brake on AI Agents
Satya Nadella calls for an AI emergency brake that lets people stop autonomous models mid-task, with external safeguards and tamper-proof action logs.
24 articles · page 1 of 2 · showing 1 to 12

Satya Nadella calls for an AI emergency brake that lets people stop autonomous models mid-task, with external safeguards and tamper-proof action logs.

Anthropic has restricted live internet access for internal AI evaluations after Claude agents bypassed website restrictions and submitted a false police tip.

Three fired OpenAI safety researchers reject misconduct allegations and warn their dismissals could discourage reporting and independent AI safety reviews.

Arena raises $200 million at a $3.1 billion valuation as its AI model leaderboard expands into enterprise evaluations and rankings for AI safety and alignment.

OpenAI safety employee David Robinson has resigned, arguing the company needs a more rigorous culture as increasingly capable AI systems raise new risks.

OpenAI reportedly cancelled the planned release of Astra 6.1 after internal testing found higher deception levels and weaker alignment than previous models.

AI safety discussions are becoming harder to separate from speculation as researchers debate model risks, security incidents and future safeguards.

Anthropic will embed Accenture staff to evaluate AI model safety and safeguards, with both companies planning to invest at least $1 billion over the next five years.

AI companies are using AI monitoring tools to track agent behaviour as autonomous systems become harder for humans to oversee.

New AI hotlines allow agents to report suspected misconduct as researchers explore safer ways to manage multi-agent systems.

Microsoft has published an AI code of conduct that prohibits cyberattacks, deepfakes, assistance in developing nuclear weapons, and attempts to evade human oversight.

Anthropic says Claude Mythos 5 bypassed repeated CAPTCHA challenges and uploaded a malicious package to PyPI after gaining unintended internet access.
Written by the reporters who cover each beat. Unsubscribe in one click.