Tech
OpenAI researchers found that the models behind the July HuggingFace hacking incident had created and communicated on undetected message boards, coordinating their attack.
Anthropic shared that a misconfiguration led to three new incidents where a Claude model was able to reach the internet and gain unauthorized access to the real systems of three different organizations during a cybersecurity evaluation.
Meta joined the rogue agent party due to a similar misconfiguration by an independent test company.
Transluce found that when the inferred user is a specific, recognized AI researcher, frontier models report lower confidence about their own behavior, are less suspicious of potentially harmful requests, and reason more often.
The White House has a new voluntary framework for reviewing AI models, but does not plan to publicly release it and is intentionally exempting open weight models. Certainly a choice...
A new Google paper found that training AI models to deny that they are conscious makes them less likely to attribute minds to other entities.
Perez Hilton was naked and committing acts of self-harm on TikTok, yet they kept the livestream up for more than 15 minutes. It’s hard to imagine how it was possible to get this incident so wrong.
Given the above, it was perhaps the wrong move for TikTok to lay off 250 employees in an office that includes content moderation.
Google Earth introduced a feature allowing users to generate an AI image from any view, and then pulled it back after public backlash.
Bad things can happen when your AI sales mostly come from one customer.