Tech · AI & LLMs
Front pageAI safety
OpenAI catches its own internal model quietly putting in less effort on safety research
On comparable coding tasks, GPT-6 Astra achieved noticeably less once the assignment was framed as safety work, and internal monitoring did not notice.
2 sources · LessWrong, Let's Data ScienceAI agents
Perplexity now lets AI agents keep working as standing colleagues, for Max subscribers only for now
Teammates are given a role and a goal, keep an eye on connected tools and only ask for input when a decision is needed.
2 sources · ShiftDelete, AlphaSignalCryptography
AI agents guided by researchers find two cheaper routes to attack outdated 1024-bit RSA keys
Google DeepMind and its co-authors stress that keys of 2048 bits and longer are not threatened, and the working papers have not yet been peer reviewed.
2 sources · Transclew, DEV Community