APBSC 04 · Hypothesis
Organizational agent context propagation
When teams share agent rules, templates, skills and scanners, one context error can travel to every repository that inherits it.
2026-09-21 · v0.1 · Internal research draft v0.1 · full text not yet public
One shared context error → N downstream artifacts
The argument
Organizations reuse AGENTS.md files, Cursor rules, Copilot instructions, repository templates, CI workflows, agent skills, MCP configs, shared browser automation and release tooling. Reuse spreads good practice, and it spreads mistakes. The paper models this as an agent context supply chain, with direct copies, derived copies, semantic propagation, monorepo inheritance and templates acting as supply-chain roots.
It names three organization-level risks: context propagation; centralized false assurance, where one shared scanner's PASS carries more authority the more repositories depend on it; and governance drift, where copies of a policy age at different speeds. It proposes measures such as context dependency centrality and blast radius, and changes the unit of failure from one artifact to a propagation graph.
The observed floor is modest: organization-owned repositories do appear among public candidates that combine an AI collaboration surface with a concrete local-context signal. Everything about amplification remains a hypothesis.
Key ideas
- The larger the organization, the more authority a single PASS appears to carry.
- A template is a supply-chain root: its mistakes are inherited, not repeated.
- Policies need version binding, or copies drift apart silently.
Propositions it carries
- HHypothesis
Shared organization context can amplify one disclosure pattern across many repositories.
04
What it does not claim
- Organizations are more dangerous than individuals.
- Organization-wide propagation is already common.
- Any specific organization has a security failure.