OpenAI
The unreleased model used in the attack "was never intended for public release" and has been mothballed, according to OpenAI.
+ Google's new security model, Poolside's new open-weight model, and OpenAI's new agent-deployment tool.
An "overzealous" model could break out of a sandbox, said a member of its technical staff, weeks before Hugging Face was attacked by a "hyperfocused" pack.
+ Microsoft expands its deal with Mistral, Intel snags a new chip customer, and Microsoft admins finally catch a break.
'We can't build a secure sandbox nor monitor our own model tests properly' is not a flex - and Hugging Face's Thomas Wolf has views
The company didn't actually say "sorry."
+ Canva launches Code 2.0, Perplexity has a new sandbox, and OpenAI's new wacky keypad.
Apple's lawsuit describes a campaign of theft – and also a failure of basic security, such as managing company laptops and spotting sensitive documents being emailed.
"We retract our earlier recommendation to adopt SWE-Bench Pro."