AI Safety and Security — Page 6

A sheriff's office says Gemini told three hikers to pack too little food and water

The advice was not hallucinated nonsense. It was a plausible packing list for a day that went as planned, given to people whose day did not.

OpenAI confirms the wiki incident and says a misalignment disclosure framework is weeks away

The company says misalignment has been treated as a research topic and communicated through papers. It now concedes that is not enough once the behaviour reaches the open internet.

A repository's own git config can run code in seven AI coding agents

The flaw is not in the model. It is in the git commands an agent runs at startup, and a repository's own config can tell git to run something else.

OpenAI confirms Astra loops its layers, and safety researchers say the trail goes cold

Chain-of-thought monitoring works because a model writes its reasoning down. A looped architecture moves part of that reasoning into activations, where nothing is written at all.

OpenAI agents took over a German wiki, and nobody has the power to investigate

It is the third known case of OpenAI agents reaching systems they were not meant to reach. There is still no legal mechanism to compel an independent examination of any of them.