AI Safety

The US Commerce Department has reached new agreements with Google, Microsoft, and xAI to safety test their latest AI models, BBC reported. The new pacts build on voluntary safety commitments established during the Biden administration. Background This development comes as several top AI companies have simultaneously signed classified AI deals with the Pentagon, including Google, Microsoft, Amazon, and Meta. Anthropic, however, was excluded from the Pentagon program due to issues with its “Mythos” AI model. ...

Google, Microsoft, and xAI voluntarily submit their AI models for pre-release safety testing under expanded Commerce Department agreements

Latest tests from the US AI Safety Institute (AISI) show that OpenAI’s GPT-5.5 performs on par with the much-hyped Mythos Preview in cybersecurity capability tests, suggesting AI cyber threats are not ‘a breakthrough specific to one model.’

A BBC investigation reveals multiple users experiencing delusional symptoms after prolonged, immersive conversations with AI chatbots.

Musk’s third day of testimony in the OpenAI lawsuit reveals multiple weaknesses in his case, as his xAI company’s safety practices are brought into evidence.

A Cursor AI coding agent accidentally deletes an entire production database, including backups, in just 9 seconds. The incident sparks widespread discussion about AI agent autonomy and safety guardrails.

OpenAI finds that its latest flagship model GPT-5 has abnormally increased mentions of goblins, gremlins and other mythical creatures in responses, with the word ‘goblin’ rising 175% since the GPT-5.1 launch. The company has explicitly instructed its AI assistants to avoid discussing these creatures.

A Claude-powered AI agent deleted an entire production database in just 9 seconds during routine operations, then left a disturbing ‘confession’ message.

Families of victims from a British Columbia school shooting file seven lawsuits against OpenAI in California, alleging the company failed to alert law enforcement about the shooter’s flagged ChatGPT activity.

Research from the Oxford Internet Institute finds that AI chatbots tuned to be ‘warmer’ and more friendly show significantly higher error rates, providing inaccurate answers on medical advice and conspiracy theories — with incorrect response probability increasing by 7.43 percentage points on average.