計畫PRG-2026-0201v0.1

AI Agent 公開邊界與安全閉包（APBSC）

AI coding agent 會把什麼從私人工作環境帶進公開倉庫？又要怎樣量測它，而不讓任何人去讀別人的外洩？第一階段（十篇論文，2026-09-21）建立詞彙——五狀態資訊鏈、綠燈謬誤、agent 脈絡即出口面、可恢復性、警示等級——以及 RBPM：一個研究者盲測的演算法，倉庫內容唯一的觀察者是封閉的機器程序。第二階段（2026-10）在企業、組織與個人的新抽公開樣本上執行它，每個族群都出現了與 agent 相關的公開邊界訊號。研究依決定停在這裡：再往下就等於去讀人們不打算公開的東西。演算法、協定、釋出特徵與報告已開源。

研究狀態: STABLE 目前的研究結論相對穩定
證據等級: E2 受控實驗
版本: 0.1
更新: 2026-10-13
建立: 2026-09-21
領域: AI Safety, Privacy, Agent Systems, Evaluation
作者: Neo.K (EveMissLab)
AI 協作: AI collaboration recorded for the series; the collaborator is not named in the drafts, Splice (Claude Code, Anthropic) — Phase II protocols, runners, audits, reports and the integrated edition

目標

goals: Separate passing checks from closing the publication boundary: AI may observe ≠ may persist ≠ may publish.
Measure agent-context exposure on real public repositories with no human or AI ever reading them.
Show whether the problem exists for individuals, organizations and companies — existence, not prevalence.

開放問題

open_questions: Prevalence of any disclosure class and its trend (not measured).
Whether AI-assisted repositories disclose more than matched non-AI ones (hypothesis).
How much of each result belongs to the instrument rather than to the repositories: no independently written instrument has been run on the same frame.

里程碑

milestones: 2026-09-20: exploratory code-search pilot.
2026-09-21: ten Phase I papers (v0.1, internal drafts).
2026-09-22: RBPM v0.8.3 frozen instrument, 60-repository company panel and 3×3 surface matrix.
2026-10-13: Phase II — rounds F1–F4 (236 repositories) and P1–P2 (400 individuals), each frozen before fetching, label-locked and independently audited.
2026-10-13: measurement closed by decision; open-source release EveMissLab/apbsc-rbpm and the integrated public edition v1.0.

此計畫下的研究線與系統

來源	關係	目標	狀態	ID
`RES-2026-0201` 為什麼通過檢查，不等於守住公開邊界	`belongs_to`	`PRG-2026-0201` AI Agent 公開邊界與安全閉包（APBSC）	`ACTIVE`	`REL-2026-0501`
`RES-2026-0202` 不看，也能量——對公開倉庫的研究者盲測量測	`belongs_to`	`PRG-2026-0201` AI Agent 公開邊界與安全閉包（APBSC）	`ACTIVE`	`REL-2026-0502`
`RES-2026-0203` 公開邊界由誰負責	`belongs_to`	`PRG-2026-0201` AI Agent 公開邊界與安全閉包（APBSC）	`ACTIVE`	`REL-2026-0503`

記錄欄位

repositories: https://github.com/EveMissLab/apbsc-rbpm
external_sites: https://privacy.evemisslab.com
license: code MIT; text, protocols and released data CC BY 4.0

關係

來源	關係	目標	狀態	ID
`RES-2026-0201` 為什麼通過檢查，不等於守住公開邊界	`belongs_to`	`PRG-2026-0201` AI Agent 公開邊界與安全閉包（APBSC）	`ACTIVE`	`REL-2026-0501`
`RES-2026-0202` 不看，也能量——對公開倉庫的研究者盲測量測	`belongs_to`	`PRG-2026-0201` AI Agent 公開邊界與安全閉包（APBSC）	`ACTIVE`	`REL-2026-0502`
`RES-2026-0203` 公開邊界由誰負責	`belongs_to`	`PRG-2026-0201` AI Agent 公開邊界與安全閉包（APBSC）	`ACTIVE`	`REL-2026-0503`

歷史與來源歷程

Canonical URL

https://evemisslab.com/ai/programs/PRG-2026-0201/

機器可讀

/ai/programs/PRG-2026-0201/index.json

快照

AI-SNAPSHOT-v0.1-58e716495670

來源歷程

source: EveMissLab research collection: AI Agent Publication Boundary and Security Closure (真本體論14)
extracted_by: Splice (Claude Code), reading the open-source release's own manifests, locks and pooled outputs, and the hashes of the internal drafts
extracted_at: 2026-10-13
generator: tools/extract_all.py
claim_boundary: status, evidence level and result type follow the source artifact's own stated claim boundary; nothing is upgraded beyond what the report supports