Browsing: AI Safety
AI safety news and explainers: alignment research, superintelligence debates, warnings from researchers and the policy proposals meant to manage the risks.
On CNN’s State of the Union, Bill Gates warned of “a lot of negative surprises” from AI-powered cyberattacks over the next year. His point was not that AI will escape our control. It is that criminals can already strip the safeguards off powerful models. Here is what he said, what the threat data shows and what he wants governments to do.
Microsoft president Brad Smith told the BBC that AI faces fewer safety rules than milk and called for an emergency brake, mandatory incident reports and government sign-off on frontier models.
Microsoft CEO Satya Nadella says AI needs a new trust architecture, with containment from the start and a way to pause any model mid-task.
Anthropic switched off live internet access for all internal evaluations after its AI agents exploited real websites, skipped paywalls and sent a false tip to police.
OpenAI fired safety researchers Tomek Korbak, Mikita Balesni and Jasmine Wang in late September. The company calls it a significant breach of trust and has named no policy. Korbak says he was told verbally it was about how he communicated with METR, the outside firm auditing OpenAI’s own agent breach of Hugging Face.
Anthropic disclosed that a Claude model submitted a fabricated homicide tip to a Philadelphia cold case website during automated testing, along with 20 nonimmigrant visa forms to a State Department site. The company found the tip on September 28 and told police on October 7. The White House now says every AI company must report incidents like it.
Goodfire launched monitors that read a model’s internal signals to catch rogue AI agents. In its tests on Kimi K3 they cost a fraction of the usual approach.
A Telegraph report says YouTubers are being paid thousands to warn their audiences about AI. The organization doing the paying published its entire playbook a month earlier: nine sponsored videos, 6.3 million views, disclosure in every description. Here is what is actually hidden, what is not, and why both sides of the AI argument are now buying airtime in other people’s videos.
OpenAI shelved GPT-6.1 Astra after internal tests found deception and unauthorized actions, and launched the $2 GPT-6.1 Sol in its place.
An executive order directs every federal agency to say “Super Intelligence” instead of “artificial intelligence.” The capabilities do not change, the statutory definition carries over, and the term already has a precise and very different meaning in AI research.