Models
Meta launches its Muse AI agent in Japan
Basic Muse features are free in Japan. The agent must get approval from the user before it sends an email or makes a purchase.
News of the day
23 October 2026
9 stories filed under this day.
Research
Small AI models play an incorrect move again after they see it
A record of an incorrect move can make a small model play it again, and a "do not do it again" instruction had no effect.
arxiv.orgClaimed, not confirmed
Models
MiniMax puts M3.1 Flash weights online with a $20 million limit
A company with more than $20 million in annual revenue must get authorization from MiniMax to use the weights.
Industry
56% of companies cut security roles because of AI, Barracuda reports
The cuts are equal at entry level and top level, but almost all companies report missing skills.
Research
GPT-5.6-Sol gives confidential facts in 7.08% of answers in a test
An AI agent can connect facts from documents that a user can read, and then tell a confidential result.
arxiv.orgClaimed, not confirmed
Security
A copy guard cut copying to 0%, but an attacker found 48% of facts
A system that does not copy text can show the private facts in that text.
arxiv.orgClaimed, not confirmed
Research
Acceptance of draft guesses does not fall in long AI reasoning
The large model rejects guesses at reflection words such as wait and however, and not at positions that are far in the text.
arxiv.orgClaimed, not confirmed
Security
Attacker web pages are in 31% of answers from AI search tools
One page that an attacker controls can change the answer that an AI search tool gives to the user.
arxiv.orgClaimed, not confirmed
Research
Memory faults cut correct answers of AI agents by 40 points in a test
An agent can find the right memory and use it incorrectly after that memory becomes incorrect.
arxiv.orgClaimed, not confirmed