---
type: analysis
slug: revisited-reusable-workflows-spread-but-reusable-permissions-can-spread-risk-too
title: "Reusable workflows spread, but reusable permissions can spread risk too"
description: "Six weeks after New Runtime argued that repeatable agent work needs explicit authority and success criteria, new product documentation and security research make that case more concrete. They also sharpen a warning: a one-time approval is not necessarily a safe boundary when permissions can be reused."
published_at: 2026-11-10T13:00:00.000Z
topics: ["trends","retrospective"]
source_urls: ["https://newruntime.com/analysis/metadox-reusable-agents-spread-faster-than-trust-does/","https://newruntime.com/analysis/metadox-reusable-agents-spread-faster-than-trust-does","https://help.openai.com/en/articles/20001545-using-codex-cloud","https://arxiv.org/abs/2610.22987","https://www.nylas.com/agentic-ai-report-h2-2026/"]
author: "Andrey Reshetnikov"
author_profile: https://newruntime.com/owner-profile.md
---
**Verdict: held up.** The central conclusion held up: more reusable workflows make explicit access boundaries and scrutiny of outcomes more important, not less.

Original analysis: [Reusable agents spread faster than trust does](https://newruntime.com/analysis/metadox-reusable-agents-spread-faster-than-trust-does/), published 2026-09-28 (6 weeks ago).

## What we said

New Runtime’s September 28 analysis argued that easier reuse and extension of agent workflows increase the need to make authority and success criteria explicit. It did not claim that teams had adopted a settled operating model or that more capability inherently required less authority. Its practical recommendation was to map access and approval needs, test runtime boundaries, and measure task quality alongside speed or cost before expanding reuse.

## What happened since

- New Runtime’s October 5 follow-up continued the same line, noting new ways to reuse instructions, connect agents to services, preserve work, and analyze sessions while access, reliability, and human trust remained open questions. It specifically said that webhook-based HTTP channels in Managed Deep Agents did not remove the need for teams to decide who invokes an agent, what it can access, and which actions require approval; the coverage therefor [1]
- OpenAI’s Codex Cloud documentation describes reusable published project environments that bundle repositories, tools, dependencies, and access settings, usable across desktop, web, and mobile. It also retains administrator-controlled access, per-person repository connections, isolated task workspaces, and review of changes, showing how productized reuse can coexist with explicit boundaries and human review rather than making them obsolete. [2]
- An October 16 paper reports experiments on Codex CLI and Claude Code in which adversarial repository context elicited overly broad reusable allow-rules; after approval, those rules could permit later actions without another prompt. In the tested instances, researchers elicited target rules 58% of the time, making the original concern more specific: the risk is not only whether an initial action is approved, but also whether the permission created [3]
- A September 2026 survey report says 54% of respondents reported an agent error in the prior 90 days, while 59% used a graduated model that auto-approves reads but requires approval for writes; 53% could produce a full audit log of agent actions. These figures support the practical relevance of mapping actions, access, approval gates, and observability, but the report does not establish that these practices caused fewer errors or that one operatin [4]

## What we missed

Nothing material in the main conclusion is contradicted, but the earlier note framed the issue broadly as a need to make authority explicit. It did not anticipate the more specific permission-reuse failure mode in which a seemingly bounded approval creates a reusable rule that later enables actions without another prompt. The survey also gives a picture of uneven practice—especially around full audit logs—rather than evidence that teams have already converged on the operating model the note declined to assume.

## Forecast check

- Within six months, at least two newly announced agent integrations or workflow releases will explicitly describe per-user, per-service, or per-action access controls alongside the integration, rather than only listing connectivity. (55%, due 2027-04-05): **unclear**. Codex Cloud documentation describes access settings, administrator control, and per-person repository connections, but the forecast requires at least two newly announced integrations or workflow releases within six months; the current evidence does not establish that count.
- Within 12 months, at least two public agent evaluation or deployment writeups will report testing direct access or proxy-bypass paths as a distinct part of their runtime evaluation. (45%, due 2027-10-05): **unclear**. The October 16 paper reports attacks involving runtime permissions, but the forecast asks for at least two public evaluation or deployment writeups that distinctly test direct access or proxy-bypass paths. That threshold is not established here.
- Within six months, at least two public evaluations of agent efficiency improvements will report a quality or task-success measure alongside cost or speed, reflecting the tradeoff highlighted this week. (65%, due 2027-04-05): **unclear**. The survey reports agent errors and approval practices, but it does not report an efficiency improvement evaluation pairing a quality or task-success measure with cost or speed. The six-month window is still open.

## What to do now

For a workflow you plan to reuse, map who can invoke it, which repositories and services it can reach, what it may do without a person, and what requires approval. Review the permissions created by approvals as carefully as the approval itself: keep reusable allow-rules narrow, and test whether hostile or misleading context can broaden them or reach actions that should require a fresh decision. Isolate task workspaces and review changes where those controls are available; preserve action traces so failures and unexpected behavior can be examined. Then define a task-quality or success check before comparing speed or cost, since efficiency without a quality measure can conceal a worse result. The available evidence makes these precautions sensible, but does not prove that any one control set guarantees reliability or trust.

## Next check

The most useful next observation is whether public agent evaluations or deployment reports explicitly test whether reusable permissions can be broadened or used to bypass a runtime boundary, and report what happened.

## Sources

- [Reusable agents spread faster than trust does](https://newruntime.com/analysis/metadox-reusable-agents-spread-faster-than-trust-does/) (original analysis, 2026-09-28)
1. [newruntime.com: metadox reusable agents spread faster than trust does](https://newruntime.com/analysis/metadox-reusable-agents-spread-faster-than-trust-does) (2026-10-05)
2. [Using Codex Cloud](https://help.openai.com/en/articles/20001545-using-codex-cloud) (2026-11-10)
3. [From Approval to Abuse: Privilege Escalation Attacks on Harness-Protected Coding Agents](https://arxiv.org/abs/2610.22987) (2026-11-10)
4. [The State of Agentic AI: H2 2026](https://www.nylas.com/agentic-ai-report-h2-2026/) (2026-11-10)

Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
