Hello, @shinichiro.matsuo, @JBringer, @Mitchell,
Following the security information sharing session, I’d like to suggest we schedule a follow-up meeting. The proposed agenda would be:
For example, would 8 - 9 AM EDT on any of the following dates work for you?
March 30, 31, April 1, 2, or 6
Please let me know your availability.
Thank you!
April 1 & 2 would work for me @Shohei_Mitani
Can we do it on April 2?
@Rola @JBringer @Shohei_Mitani
I am available from 8:00 AM EST if you would like to have the meeting on April 2.
Happy to hear everyone else’s opinion
8 AM on April 2 seems like the best option.
@shinichiro.matsuo shall I send out an invitation for tomorrow?
Yes please
Here is the Zoom meeting invitation:
BGIN Admin is inviting you to a scheduled Zoom meeting.
Topic: BGIN Security Agentic AI Project - Meeting 8
Time: Apr 2, 2026 08:00 Eastern Time (US and Canada)
Join Zoom Meeting
Meeting ID: 869 1099 8014
One tap mobile
+13017158592,86910998014# US (Washington DC)
+13052241968,86910998014# US
Join instructions
and here is the Google Calendar link:
If you would like me to create an independent thread on discourse, or share on the social media accounts, please let me know.
Thank you for the invite.
Draft Proposal for Agent Goals
(Goal for today: Align on the proposed direction and establish a timeline)
From Concept Demo to Practical Application
While the bot demo illustrated the concept, real-world deployment requires defined use cases, a robust knowledge base, and primarily a secure pipeline.
Legal Barriers & the “Nine Line” as Near-Term Fix
Overcoming legal friction requires defining a minimum emergency information set to safely share critical insights, making governance design just as decisive as model performance.
Compression, Reconstruction & Expert Pathways
LLMs allow experts to compress/abstract threat analyses into standard templates that counterparts can safely reconstruct within their own private contexts, enabling measurable cross-organizational skill portability.
KB Architecture & Governance
The core challenge is the operational infrastructure—balancing data quality, privacy boundaries, and progressive trust—proving that governance, rather than model tuning, is the primary research battleground.
Japanese-Language, Real-Time Access
LLMs bridge crucial natural-language barriers and technical stakeholder gaps, with success directly measured by localized response times, recall, and domain translation quality.
Provenance Value: Rediscovery + C-Suite Decision Support
Tracing historical incident patterns through provenance graphs directly supports both engineers in avoiding recurring losses and executives in prioritizing limited security budgets.
Standard Taxonomies: MITRE Frameworks + Crypto CVE/NVD
Applying traditional cybersecurity frameworks to crypto dramatically improves LLM accuracy, though the current absence of a standardized crypto vulnerability repository destabilizes information sourcing.
Tabletop Exercises with Assigned Personas
Running simulation exercises with assigned roles frees participants from corporate constraints, allowing practical standards to emerge organically from observable experiments rather than theoretical paper designs.
Demonstrate Clear Value: Establish and validate specific use cases where the information-sharing agent measurably improves security outcomes.
Establish Global Security Policy: Define comprehensive policies governing how the agent processes, analyzes, and stores cross-organizational shared data.
Design Boundary-Aware DB Framework: Design an architecture that cleanly separates confidential, organization-specific data (internal KG) from abstracted, shareable insights.
Implement Governance: Design automated sensitivity labeling, explicit separation rules, and dynamic privacy controls.
Maintain Information Quality: Create incentive structures and automated verification to assure continuous data reliability.
Enable Executive Decision Support: Leverage the knowledge base to provide actionable insights for C-suite vulnerability prioritization and budget allocation.
Draft Minimal Security Policy: Formulate foundational rules for the agent’s safe handling of sensitive information.
Build Secure Private Enclaves: Create environments where users can safely upload confidential data that the agent can query via private inference without sharing it externally.
Define example high-value data elements: Establish and populate the minimum viable shared information set containing high-value data elements.
Integrate External Resources: Connect the agent to standardized sources like MITRE ATT&CK/D3FEND/AADAPT and available crypto incident databases.
@JBringer
Could you join the call now?