• Home
  • News
  • Blog
  • Releases
  • LLM history
  • Compare LLMs
  • Library
  • About
⌘K
Sign in

A blog and notes on development. The easiest way to reach me is via the social links below.

Contacts
talalaev.misha@gmail.com
Documents
Personal data processing policyPersonal data processing consent
Photo: Frengki Mulia / Unsplash

A Microsoft Copilot vulnerability, Reddit AI media and new agents

Sh0ny
Sh0ny
18 августа 2026
  1. Home
  2. Blog
  3. A Microsoft Copilot vulnerability, Reddit AI media and new agents
2 min read

In short

In today's roundup: a Microsoft 365 Copilot compromise, video and podcast generation from Reddit posts, and several tools for local AI agents. Plus pieces on security and on training developers in the age of AI.

In today's roundup: a Microsoft 365 Copilot compromise, video and podcast generation from Reddit posts, and several tools for local AI agents. Plus pieces on security and on training developers in the age of AI.

🔥 Hot:

🔹 Researchers hacked Microsoft 365 Copilot and got it to disclose passwords and other data — The attack made it possible to obtain sensitive information without the user's confirmation. 🔹 Reddit turns user posts into podcasts and short videos — The platform is adding AI formats for reworking text content.

➡️ News:

🔹 Engrava stores AI agent memory in a local graph database on SQLite — The project is distributed under the MIT licence and is meant for local use. 🔹 PhysiClaw lets an AI agent physically operate an iPhone — This is an open-source agent project that interacts with the device directly. 🔹 Krystal Loop Protocol defines a bounded worker/critic loop for AI coding — The approach separates carrying out a task from having a critic agent check it. 🔹 Engrava offers local graph memory for AI agents — The database runs on top of SQLite and is available under the MIT licence.

➡️ Useful reading:

🔹 Maturity in adopting AI and maturity in securing it almost never match — The piece proposes rating both sides on the CMMI scale and ties them to FSTEC requirements and Russian standards. 🔹 AI is taking from juniors the routine tasks on which skills used to be built — The analysis shows why the disappearance of automated tests, analysis and documentation as learning practice creates a problem for training engineers. 🔹 OneShot proposes unambiguous specifications for AI coding agents — The piece is about stating tasks precisely so an agent can carry them out predictably.

➡️ Discussions and cases:

🔹 A user is looking for a compact 4B model to explain code without vibe-coding — They want a local assistant for learning to program, finding simple errors and answering without constant hallucinations. 🔹 AI Stock Research Assistant shows an example of an assistant for stock analysis — The project is published as a Streamlit app.

📝 If you would like to add other news and materials to the list, write in the comments.

новости
More AI-tool write-ups on the Telegram channel — short and to the point
Subscribe on Telegram

Comments

(0)
​