Daily Briefing

June 29, 2026

2026-06-28

30 articles

Bringing more control over your connectors

Date not detected

Summary

Mistral AI announced new security and control features related to connectors, including enhanced admin control functions, connector-wide API keys, multi-account support, and debugger and workflow support for secure integration with external enterprise platforms.

Key Points

Introducing enhanced admin control features that support setting up connector access for each workspace and activating/deactivating individual tools
Provide connector-wide API keys to prevent impersonation when linking third-party systems in automated AI tasks
Connector debugger for end-to-end root cause analysis of MCP connectors and support for multi-account authentication on a single connector

Notable Quotes & Details

Intended Audience

Enterprise system administrators, AI developers, and IT security personnel

Workflows for work that runs the business

Date not detected

Summary

Mistral AI has released a public preview of 'Workflows', an orchestration layer for stable production transition of enterprise AI processes.

Key Points

Workflows work on laptops but provide durability, observability, and fault tolerance to address issues such as silent failures or inability to withstand network timeouts in production.
Developers can write workflows in Python and publish them to Le Chat so that anyone in the organization can run them, and every step can be tracked and audited in Studio.
It can be implemented to pause the workflow and wait for human approval with a single line of code such as wait_for_input(), and is already being used in global shipping document verification and KYC processes.

Notable Quotes & Details

Notable Data / Quotes

ASML, ABANCA, CMA-CGM, France Travail, La Banque Postale, Moeve
wait_for_input()

Intended Audience

Enterprise teams looking to automate enterprise AI application developers, IT operators, and business processes

Introducing Forge

Date not detected

Summary

Mistral AI has launched Forge, a system that helps companies build customized AI models that understand their unique domain knowledge and workflow by learning from their own data, documents, and codebases.

Key Points

Forge is a frontier-level AI model building system that reflects the company’s internal proprietary knowledge, regulations, and workflow.
It supports modern learning methods throughout the model life cycle, including pre-training, post-training, and reinforcement learning.
By training and controlling models in their own infrastructure environment, companies can secure data sovereignty, intellectual property (IP), and regulatory compliance.

Notable Quotes & Details

Notable Data / Quotes

ASML
DSO National Laboratories Singapore
Ericsson
European Space Agency
Home Team Science and Technology Agency (HTX) Singapore
Reply

Intended Audience

Enterprise customers and developers who want to build custom AI models and agents based on their own security data and expertise

Mistral AI partners with NVIDIA to accelerate open frontier models

Date not detected

Summary

Mistral AI and NVIDIA are collaborating to form the NVIDIA Nemotron Alliance to accelerate the development of open frontier artificial intelligence models and disseminate joint research and models.

Key Points

Mistral AI participates as a founding member of the NVIDIA Nemotron Coalition and jointly develops frontier open source AI models.
The two companies will combine Mistral AI's specialized model architecture and full-stack platform with NVIDIA's computing resources and synthetic data generation pipeline.
Mistral AI has launched Mistral Small 4, a new open model to support developers around the world.

Notable Quotes & Details

Notable Data / Quotes

Mistral Small 4
“Open frontier models are how AI becomes a true platform,” said Arthur Mensch, cofounder and CEO of Mistral AI.
NVIDIA DGX Cloud
NVIDIA Nemotron 4

Intended Audience

AI developers, researchers, and enterprise customers

Leanstral: Open-Source foundation for trustworthy vibe-coding

Date not detected

Summary

Mistral AI has released Leanstral, the first open source code agent for Lean 4 proofreading.

Key Points

Leanstral is a highly efficient open source code agent with a 6B active parameter scale designed for going beyond simple mathematical problem solving to verifying real-world formal software repositories and specifications.
Weights are released under the Apache 2.0 license and are available through Mistral vibe's agent mode and free API endpoints.
In a benchmark (FLTEval) based on the FLT project's PR, Leanstral-120B-A6B demonstrated superior efficiency and performance over much larger large-scale open source models (GLM5, Kimi-K2.5, etc.).

Notable Quotes & Details

Notable Data / Quotes

Lean 4
6B active parameters
Apache 2.0
FLTEval
lean-lsp-mcp
Claude Opus 4.6, Sonnet 4.6, Haiku 4.5
Qwen3.5 397B-A17B, Kimi-K2.5 1T-A32B, GLM5 744B-A40B
Leanstral-120B-A6B
GLM5-744B-A40B
Kimi-K2.5-1T-32B
16.6
20.1

Intended Audience

Software verification engineers, math researchers, AI developers, and open source artificial intelligence communities.

California will tax downloaded software for the first time as part of a $351.7 billion budget deal

2026-06-28

Summary

The state of California agreed to a $351.7 billion budget plan that would impose a consumption tax on downloaded software for the first time to secure tax revenue and increase reserves to prepare for windfalls from AI company IPOs.

Key Points

California's state government and legislature have agreed on a budget proposal that would expand the consumption tax on off-the-shelf software downloaded from the web.
This software tax is expected to secure a total of $2 billion in additional tax revenue each year, including $900 million for state governments and $1.1 billion for local governments, starting in fiscal 2028.
A bill to set aside more budget in the event of a surge in profits following the IPO of California-based AI companies such as OpenAI and Anthropic is scheduled to be submitted to a referendum in November.

Notable Quotes & Details

Notable Data / Quotes

351.7B
2B
900 million
1.1 billion
2028
June 18
Most of us don’t get prewritten software on a physical disc anymore. The whole world is past that, our tax code isn’t
For millions of Californians, this isn’t abstract. This impacts real people, real businesses. This tax could be the difference between making payroll and missing it
35.2 billion
4.5 billion
90 million
250 million
3 billion
2030
5 million
70%

Intended Audience

IT business insider, software consumer, tax law and public policy analyst

The BIS warns an AI bust could hit credit markets as hard as the 2008 financial crisis

2026-06-28

Summary

The Bank for International Settlements (BIS) warned in its annual report that if the AI investment boom collapses, it could cause a major blow to the credit market on the scale of the 2008 financial crisis.

Key Points

BIS pointed out that disappointment in AI investment returns could lead to a sudden withdrawal of funding, turning the investment boom into a long-term recession and sending a knock-on shock to financial markets.
It warned of the vulnerability and opacity of the 'circular finance' structure, in which semiconductor manufacturers and hyperscalers invest equity in AI labs and then purchase chips or computing power from investors.
The concentration of AI stocks now exceeds that of the dot-com bubble, with the top 10 companies in the S&P 500 accounting for 36% to 40% of the index.

Notable Quotes & Details

Notable Data / Quotes

2008
36%
40%
Disappointment in returns could trigger a sudden pullback in financing and turn the capex boom into a protracted investment bust, with potential knock-on effects on financial conditions

Intended Audience

Financial regulators, institutional investors, economic policy makers and financial market participants

Auto repair is one of the least digitised industries in America. AI is changing the economics of why.

2026-06-28

Summary

With the introduction of AI-based software into the automobile maintenance industry, a representative analog industry in the United States, the market structure and economic feasibility are changing.

Key Points

Approximately 280,000 auto repair shops in North America still operate with the same manual work methods as in the 1990s, such as phone reservations and paper order forms, but digitization is progressing rapidly as the burden of data input is eliminated with the introduction of AI.
To solve the industry's biggest problem, an unanswered call rate of over 40%, an AI hospitality service that allows 24-hour reservations and follow-up was introduced, leading to an immediate increase in sales.
The contract value per repair shop is increasing through AI-based operations such as demand forecasting schedule management and automatic customer management, and the introduction of software is accelerating along with the activation of garage mergers and acquisitions (rollups) by private equity funds (PE).

Notable Quotes & Details

Notable Data / Quotes

280,000 US auto repair shops
$3.4 billion in 2026 to $8.6 billion by 2033
14.2% CAGR
missed-call rates above 40%

Intended Audience

Automobile maintenance industry worker, startup founder and venture capital investor in the vertical AI field

Google is rationing Gemini access to Meta because it cannot provide enough compute

2026-06-28

Summary

As Google limits the use of Meta's Gemini AI model due to a lack of computing power, Meta is accelerating the transition to its own AI model.

Key Points

Google has limited access to Gemini AI models for major customers, including Meta, due to a lack of computing capacity.
In order to reduce dependence on Gemini, Meta is moving tasks such as safety processes to ‘Muse Spark’, a new internal model developed in-house.
As the speed of AI infrastructure construction cannot keep up with consumption, physical compute bottlenecks are occurring across the industry, such as Google renting NVIDIA GPUs from SpaceX.

Notable Quotes & Details

Notable Data / Quotes

Google agreed to pay SpaceX $920 million per month for the use of 110,000 NVIDIA GPUs to meet Gemini enterprise demand.
Google is spending more than $180 billion on capital expenditures (capex) this year.
Meta presented a 2026 capital expenditure guideline of $115 to $135 billion and cut 8,000 employees in May, reallocating 7,000 to its AI department.

Intended Audience

Readership interested in AI business trends and IT infrastructure industry

India’s payments chief says AI will drive UPI from 750 million to a billion daily transactions

2026-06-28

Summary

The CEO of National Payments Corporation of India (NPCI) announced that the company will leverage AI technology to increase the daily transaction volume of the Unified Payments Interface (UPI) from 750 million to 1 billion.

Key Points

AI will drive UPI's next-generation user acquisition through fraud prevention, credit distribution, and multilingual voice onboarding.
In the Indian payment market, PhonePe and Google Pay occupy more than 80% of the market, and regulations to eliminate monopoly are being discussed.
NPCI CEO Dilip Asbe emphasized that the Indian fintech ecosystem has a huge opportunity to build precise and clear small language models (SLMs).

Notable Quotes & Details

Notable Data / Quotes

750 million
a billion
80%+
2023
1%
December 31, 2026
"AI will be used very effectively when we look at the next wave of UPI, and that includes all aspects, including reaching new users,"
"We have a very rich data set in our ecosystem, I think there is a big opportunity for Indian companies, the banks, FinTechs, and the ecosystem, to create small language models which are sharp, specific, and as deterministic as possible."

Intended Audience

Fintech industry players, financial technology investors, and IT professionals interested in AI and Indian payments market trends

Prosecutors used ChatGPT logs as evidence in the Palisades fire trial

2026-06-28

Summary

Prosecutors used the defendant's ChatGPT conversation records as evidence in the Palisades wildfire arson trial, but failed to convince the jury, so the trial was dismissed.

Key Points

The prosecution presented ChatGPT conversation records, along with iPhone location data and CCTV footage, as evidence to prove defendant Jonathan Rinderknecht's arson charges.
ChatGPT records included requests to create images of fire, questions about anger, complaints about the wealthy and questions about whether cigarettes were responsible for fires started by cigarettes.
The jury was 10 to 2 in favor of not guilty, so a mistrial was declared due to a verdict disagreement (decision), and one juror said he was angry at the prosecution's induction that dismissed the use of an ordinary chatbot as a character flaw.

Notable Quotes & Details

Notable Data / Quotes

New Year's Day 2025
10-2
“I talk to ChatGPT all the time.”

Intended Audience

General readers interested in cases of AI technology being used as legal evidence and understanding of the technology in the judicial system

Building a Stable Fable 5 Traces Workflow in Colab: Parsing Tool Calls, Auditing Data, and Training Baselines

2026-06-28

Summary

This is a tutorial on building a Colab workflow that utilizes Hugging Face's Fable 5 Traces dataset to minimize dependencies, analyze, purify, and visualize the coding agent's trace data, and train a naive Bayes model.

Key Points

Instead of heavy libraries such as datasets, scikit-learn, and scipy, Colab's dependency stability was secured by setting up light environments such as huggingface_hub, rich, and tqdm.
By directly downloading and parsing the merged JSONL file, we implemented a data audit function that normalizes tool calls, normalizes text output, and inspects potential leak patterns (secret-like patterns) such as passwords and API keys.
It exports safe no-CoT conversation data that can be used for SFT learning, and trains a naive Bayes baseline model implemented in pure Python for text classification and tool usage prediction.

Notable Quotes & Details

Notable Data / Quotes

Glint-Research/Fable-5-traces
fable5_cot_merged.jsonl
0.23.0
13.0.0
4.66.0

Intended Audience

AI developers and researchers who work with coding agent trace data and want to reliably build data pipelines and model training workflows in Colab.

Liquid AI Ships LFM2.5-230M with llama.cpp, MLX, vLLM, SGLang, and ONNX Support for On-Device Inference

2026-06-28

Summary

Liquid AI has launched its smallest model, the LFM2.5-230M, which supports a variety of tools for on-device inference and agent operations.

Key Points

Liquid AI's LFM2.5-230M is the smallest open weight text model at 230 million parameters based on the LFM2 architecture.
It runs on-device at 213 tok/s on a Galaxy S25 Ultra and 42 tok/s on a Raspberry Pi 5, and supports llama.cpp, MLX, vLLM, and more.
It outperforms larger models such as Qwen3.5-0.8B and Gemma 3 1B in instruction execution (IFEval) and data extraction benchmarks, but is not suitable for math, code generation, and creative applications.

Notable Quotes & Details

Notable Data / Quotes

230M params
213 tok/s on a Galaxy S25 Ultra
42 tok/s on a Raspberry Pi 5
IFEval: LFM2.5-230M (71.71) vs Qwen3.5-0.8B (59.94) vs Gemma 3 1B IT (63.49)
293–375 MB footprint

Intended Audience

Embedded AI/agent developer developing instruction execution and data extraction functions in mobile, robotics, and hardware edge

Show GN: ArachneControl – An open source data collection system where servers remotely control browsers to collect data.

2026-06-28

Summary

An introduction to ArachneControl, an open source self-hosted system where servers remotely control users' actual browser sessions, bypassing security and login barriers and dynamically collecting data.

Key Points

It has a Zero-Footprint design that performs collection through logged-in user browsers, avoiding targeted backend load and login blocking walls.
Collection rules are issued dynamically by the server at runtime, eliminating the need to redeploy clients.
You can create selector and action sequence recipes by clicking on elements in the WebUI, and script eval is prohibited for safety.

Notable Quotes & Details

Notable Data / Quotes

ENABLE_TUNNEL=1

Intended Audience

Web data collection system and crawler developer, developer interested in QA debugging and open source tools

DSpark: Accelerating LLM Inference Using Speculative Decoding [pdf]

2026-06-28

Summary

This article is about DSpark, a speculative decoding framework that overcomes the limitations of existing parallel drafters and accelerates LLM inference by combining quasi-autoregressive generation and reliability scheduling.

Key Points

By injecting intra-block dependencies through a quasi-autoregressive structure that combines lightweight sequential modules on a parallel backbone, we solved the problem of a sharp drop in late-stage acceptance rates.
The reliability head estimates the survival probability for each location, and the hardware-aware scheduler dynamically adjusts the verification length according to engine throughput to suppress verification waste.
When deploying DeepSeek-V4 into actual service, the generation speed per user was accelerated by 60–85% compared to MTP-1, the existing production baseline, in the same throughput environment.

Notable Quotes & Details

Notable Data / Quotes

60–85%

Intended Audience

LLM AI engineers and system developers interested in optimizing inference and streamlining infrastructure serving.

Paca - Open source project management tool for human and AI agent collaboration

2026-06-28

Summary

An introduction to Paca, an open source self-hosted project management platform that supports collaboration between humans and AI agents.

Key Points

Engage AI agents as equal members of the Scrum team rather than just chatbots, collaborating with humans on sprints and Scrumban boards.
Manage tasks, documents, and sprints with natural language commands without leaving the editor through MCP server and Claude Code.
Runs in a secure, isolated sandbox environment using OpenHands SDK-based agents and WebAssembly plugins

Notable Quotes & Details

Notable Data / Quotes

P-A-C-A cycle (Plan → Act → Check → Adapt)

Intended Audience

Developers, project managers, DevOps engineers, and IT community members interested in AI collaboration tools

A peek inside Reddit anti-spam

2026-06-28

Summary

This is an analysis case in which Reddit's temporary system error in 2021 exposed anti-spam internal removal reasons and operational data that should have been treated as private in external apps.

Key Points

The code processing path of the Relay for Reddit app overlapped with the Reddit API error, exposing private anti-spam internal data.
Exposed spam blocking reasons include domain blocking, spammit probability score, shadowban details, and spamurai system.
The Reddit anti-spam system is understood to be a multi-generational system that combines Python 2.7-based inspection, Lua rules, Snooron, an image classification tool, and the Google Perspective API.

Notable Quotes & Details

Notable Data / Quotes

2021
39.71%
98.19%
2017
2026

Intended Audience

IT security and web service developers, developers interested in designing anti-spam systems

NLnet Labs' LLM Use Policy

2026-06-28

Summary

NLnet Labs has introduced a policy that strictly limits the use of LLM in project contributions and communications and emphasizes verification and accountability of human developers.

Key Points

Code and documentation contributions must be hand-written by humans and cannot include content generated by LLM or other probabilistic tools.
When interacting with NLnet Labs, including by reporting issues, vulnerabilities, or posting to community forums, you must transparently disclose whether you use LLM.
The main reason for the rule is to prevent the burden of reviewing and long-term maintenance of generated code from being passed on to the team and to protect developers' time.

Notable Quotes & Details

Notable Data / Quotes

sep@nlnetlabs.nl
10,000 lines of code

Intended Audience

Error 500 (Server Error)!!1500.That’s an error.There was an error. Please try again later.That’s all we know.

NagaTranslate: Building a translation and voice pipeline for low-resource Nagaland creoles (Whisper, VITS, LLMs) [P]

2026-06-28

Summary

Development of a translation and speech processing pipeline for the under-resourced languages of Nagaland, India - Nagamiz, Ao, and Sema - and a request for advice on technical limitations.

Key Points

We initially used a fine-tuned NLLB model for text translation, but are now leveraging the commercial LLM API and few-shot prompting for natural conversation flow.
For speech synthesis (TTS), the VITS model is fine-tuned with our own Nagamizu speech data, and for speech recognition (ASR), the Whisper model is fine-tuned and hosted on Hugging Face Spaces ZeroGPU.
Handling non-standardized spelling variations, ensuring robustness in regional dialect and accent recognition in small datasets, and moving from commercial APIs to low-cost self-hosted open weight models are key challenges.

Notable Quotes & Details

Notable Data / Quotes

Nagamese, Ao, and Sema
No Language Left Behind
Hugging Face Spaces ZeroGPU

Intended Audience

AI researchers and developers interested in developing low-resource language processing (NLP), speech recognition, and synthesis technologies

I shrank a transformer until every number fitted on the screen and made the weights editable [R]

2026-06-28

Summary

To understand how LLM works, we developed a transformer visualization webpage that displays all the figures on the screen and scales them down so that the weights can be modified directly.

Key Points

Visualizes the entire process of predicting the next word by inputting 4 words through an ultra-small transformer model with 6 vocabulary sizes and 3-dimensional embedding dimensions.
Weights and word vectors can be edited directly, and when modified, all downstream calculation results are recalculated in real time.
Implemented as a single HTML file without external libraries or build steps, and the backpropagation (training) process is omitted.

Notable Quotes & Details

Notable Data / Quotes

https://dgochin.github.io/transformer/

Intended Audience

Developers and learners who want to understand the working principles of mathematical/matrix operations of Transformers and LLM from the ground up.

We have Mythos at Home: GLM 5.2 beats Claude in our Cyber Benchmarks

2026-06-28

Summary

Zhipu AI's open weight model, GLM 5.2, showed superior performance, beating Claude Code in Semgrep's IDOR benchmark test.

Key Points

Zhipu AI's GLM 5.2 model achieved an F1 score of 39% in the IDOR detection test, outperforming Claude Code (32%).
This is the result of an experiment to determine the role of the capabilities of the model itself and the harness (scaffolding) surrounding it in vulnerability detection performance.
GLM 5.2 was released on June 13, 2026, and is distributed as Openweight under the MIT license, giving security teams the advantage of being able to run and fine-tune it on their own hardware.

Notable Quotes & Details

Notable Data / Quotes

39% F1 on IDOR detection
Claude Code (32%)
roughly $0.17 per vulnerability found
Semgrep's multimodal pipeline (53–61% F1)
June 13, 2026
June 16

Intended Audience

Security researchers, AI model evaluators, security tool developers, and the IT community

Use Android Auto? How to limit what information Gemini learns about you

2026-06-28

Summary

We will guide you through Gemini's personal information collection method integrated into Android Auto and how to set personal information protection settings to limit this.

Key Points

It's not possible to completely disable Gemini in Android Auto, but you can limit the information it collects through some settings.
If you are concerned about voice information being leaked due to the always-on standby microphone, you can turn off the always-listening function in the settings and change it to activate Gemini only with the steering wheel button.
You can individually select and restrict access permissions for call logs, text messages, contacts, etc. in the Android Auto permissions menu in your smartphone settings.

Notable Quotes & Details

Intended Audience

General users concerned about personal information collection by Android Auto and Gemini

Domestic AI industry ignores Manus... “There are abundant substitutes rather than rejection of Chinese products.”

2026-06-28

Summary

The domestic AI industry is reacting negatively to the introduction of the Chinese AI agent service ‘Manus’ due to security concerns and abundant substitutes.

Key Points

Domestic AI companies do not feel the need to introduce Manus as existing American services such as Claude and ChatGPT are dominating the market.
Technical limitations such as data leakage concerns, unstable systems, and rapid credit consumption unique to Chinese services are pointed out.
Thanks to its intuitive interface and visualized agent operation process, it received positive reviews only for some tasks, such as market research for non-developer occupations.

Notable Quotes & Details

Notable Data / Quotes

About 10 places
Internal testing was conducted as it received attention as the second Deep Seek, but there were already many highly usable alternatives such as Claude, so it was not possible to find a clear differentiation of Manus.

Intended Audience

General readers interested in domestic AI industry insiders and technology trends

Benchmarks for long-term computer use revealed... ‘Opus 4.8’ highest score

2026-06-28

Summary

AI research institute

Key Points

OS World 2.0, a new benchmark that evaluates real computer task performance over long periods of time, has been released.
It is based on a total of 108 tasks that took an experienced person an average of 1.6 hours, with Claude Opus 4.8 scoring the highest with a completion rate of 20.6%.
AI agents showed limitations when performing complex long-term tasks, such as forgetting constraints, missing intermediate information, and skipping verification.

Notable Quotes & Details

Notable Data / Quotes

26th (local time)
Average 1.6 hours
108 tasks
7 specializations and 21 sub-disciplines
Average 27.25 checkpoints
Average of 318 tool calls
Based on work budget of 500 steps
20.6%
54.8%
About 13%

Intended Audience

AI researchers, AI agent developers, and technology industry workers

DeepSeek releases open source ‘D Spark’ that increases LLM inference speed by up to 85%

2026-06-28

Summary

DeepSeek has unveiled DSpark, an open source speculative decoding framework that significantly improves LLM inference performance.

Key Points

We introduced a semi-autoregressive structure that combines the high speed of the parallel method with the high context consistency of the autoregressive method.
A reliability-based verification function has been added that automatically adjusts the number of tokens to be verified at once depending on GPU load.
By fixing the weights of the target model and learning only some components, such as the draft model, performance was improved while maintaining quality.

Notable Quotes & Details

Notable Data / Quotes

26th
85%
DSpark
DeepSpec
Creation speed per user improved by 60-85% for Flash models and 57-78% for Pro models

Intended Audience

Artificial intelligence model developer and LLM service optimization engineer

Epoch AI reveals ‘long-term development ability’ benchmark… “I can’t pass the code memorization test”

2026-06-28

Summary

Epoch AI has released a benchmark 'Mirror Code' that allows re-implementation of the entire program without the Internet or original source code to evaluate the long-term software development capabilities of AI models.

Key Points

Unlike traditional short-term task evaluations, Mirror Code measures long-term software development ability to reimplement an entire program from scratch without original code.
It provides a sufficient inference budget, and among the latest models, Claude Opus 4.7 showed rapid performance improvement, recording a success rate of 56%, but no model solved all tasks.
Among the evaluation framework and 25 tasks, 22 programs are released as open source through GitHub, and 3 are maintained as private test sets.

Notable Quotes & Details

Notable Data / Quotes

26th (local time)
25 programs
19th
2600 dollars
Claude Opus 4.7
16,000 lines of Go code
40+ commands
Gotree
2~17 weeks
14 hours
$251 (about 386,000 won)
56%
More than 90%
30%
GPT-5.5
GPT-5
Claude Opus 4.1
6 programming languages
132 tasks
Error 500 (Server Error)!!1500.That’s an error.There was an error. Please try again later.That’s all we know.

Intended Audience

AI researcher, software engineer, AI model developer

KAIST, quantitative analysis of AI's 'digital age discrimination'..."Clandestine age bias in generative AI"

2026-06-28

Summary

KAIST researchers quantitatively analyzed and identified age stereotypes and biases toward the elderly embedded in the answers of the generated AI (GPT-4o).

Key Points

As a result of analyzing the text generated by GPT-4o using the stereotype content model of social psychology, the elderly group aged 60 or older showed high 'warmth' scores but low 'competence' scores.
Uniform descriptions were repeated for people in their 70s or older, and the frequency of expressions of 'assertiveness', which indicates confidence and initiative, decreased as age increased.
If such biased depictions of AI are continuously exposed, there is a risk that it will strengthen social prejudice against the elderly and lead to digital age discrimination.

Notable Quotes & Details

Notable Data / Quotes

GPT-4o
28th
From 10 to 90 years old
900
Professor Choi Moon-jeong: “AI bias is not a technology problem, but a social problem, and for inclusive artificial intelligence, various generations must participate in the development process.”
The Gerontologist February 2026 Special Issue

Intended Audience

AI developers, social science researchers, technology policymakers, and the public interested in generative AI bias issues

SGA Solutions Red Castle acquired CC certification EAL4

2026-06-28

Summary

SGA Solutions' server security solution, Red Castle V6.0, has acquired CC certification EAL4 grade, strengthening its targeting of the public, financial, and large enterprise security markets.

Key Points

SGA Solutions' 'Red Castle V6.0 for Windows Server 2025 R3' acquired the international standard CC certification EAL4 grade.
This product blocks lateral movement attacks through OS kernel-level access control and microsegmentation-based internal network access control.
Through this certification, SGA Solutions strengthens the national network security system and public markets responding to zero trust, and seeks to expand into the financial and large corporate markets.

Notable Quotes & Details

Notable Data / Quotes

RedCastle V6.0 for Windows Server 2025 R3
EAL4
National Security Requirements V3.0
Windows Server 2025
“As the value of data increases in AI and cloud-centered IT environments, server-level security control and microsegmentation are becoming essential rather than optional.”

Intended Audience

Security managers and IT infrastructure operators at public institutions, financial institutions, and large corporations

Korea Data Industry Promotion Agency “Supporting small and medium-sized businesses with data”

2026-06-28

Summary

In order to support the use of data by small and medium-sized enterprises and small business owners, the Korea Data Industry Promotion Agency is promoting the advancement of the 'AI·Data Problem Solving Bank' service that combines generative AI and AI agent technology and will launch it around December.

Key Points

By combining generative AI and AI agent technology, we implement an interactive analysis service that supports everything from diagnosing problems to deriving implementation strategies just by asking natural language questions or uploading data.
The internal DB, recipe repository, and external system are linked in real time based on the MCP (Model Context Protocol) standard protocol and operate in an orchestrator-based structure.
The service will be reorganized in 2026 to transform into an executable AX service that allows users to receive intuitive analysis results and visualization data without complex analysis tools.

Notable Quotes & Details

Notable Data / Quotes

2026 AI/Data Problem Solving Bank Operation/Function Advancement
Service around December
2,460 use cases and 1,248 data recipes provided on the portal
This year, we plan to secure an additional 900 use cases and 1,200 AI/data recipes.
Promotion of advancement until November of this year
Yang Jae-su, President of Korea Data Industry Promotion Agency: “Even small and medium-sized businesses and small business owners without complex statistical knowledge can easily secure the necessary decision-making information by following a proven analysis flow.”

Intended Audience

Small and medium-sized businesses, startups, and small business owners who have difficulty utilizing data and resolving business issues

Space

2026-06-28

Summary

Space

Key Points

The rumor about Starlink entering the independent mobile carrier was highlighted when SpaceX COO Gine Shortwell mentioned it to investors at the IPO road show.
Space
Some are raising the possibility that it is a business in the form of a portable router that connects to smartphones via Bluetooth and Wi-Fi, or as a bargaining chip, rather than building an independent terrestrial network.

Notable Quotes & Details

Notable Data / Quotes

AWS-3
Starlink Mini
paper tiger

Intended Audience

Mobile communication and space tech industry officials and SpaceX investors

PreviousDaily Briefing

NextDaily Briefing