AlphaMaven AlphaMaven

Shared PCs for people and AI call for access checks on every request

Reported by TokenPost
Nov 7, 2026· 2 min read
Save Post LinkedIn Email
Key Takeaways
  • • Andreessen Horowitz's "Defense in Depth, Rebuilt for Machine Speed" report, published November 6, 2026, argues that a single login can no longer establish trust for requests from a shared device used by both people and AI agents.
  • • The report recommends that identity, endpoint security and network access be combined into a single access decision, made on the device where the originating principal can still be distinguished.
  • • Access should be granted only when a task requires it, with credentials scoped to one task and one resource and lifetimes measured in minutes, then withdrawn when the task ends.
  • • Actions that move money, delete data or grant access should require human approval, and every agent action should be logged against the person who launched it.
  • • Andreessen Horowitz named three measures for buyers to assess these systems: standing credentials with production access, attribution of activity to a human or software principal, and the time needed to revoke a flagged session.
Summary
According to TokenPost's coverage of an Andreessen Horowitz report titled "Defense in Depth, Rebuilt for Machine Speed," published November 6, 2026, the traditional security model of trusting a session after login is no longer adequate when a single laptop hosts a person, a coding agent with its own keys, and a browser agent with separate sessions, all sending requests in the same minute. The firm argues that trust must depend on whether a person or software produced each request, and that this determination must be made on the device, since the distinction is lost once traffic reaches the network.

The proposed architecture centers on granting access only when a task requires it. Context would determine permissions, and those permissions would be withdrawn when the task ends. The report attributes current friction to standing privileges that require periodic review, and contends that removing them would allow most access decisions to occur in the background. Under the model, an agent on the device would identify the process and the human or software principal behind each outbound request and cryptographically sign it. A broker would issue task-specific credentials, and the network would reject requests that are unsigned or fall outside their permitted scope.

Human oversight remains central in the framework. Actions that move money, delete data or grant access would require explicit human approval, and every action an agent takes would be logged against the person who launched it, preserving accountability for automated work.

For buyers evaluating such systems, Andreessen Horowitz identified three measures: the standing credentials that retain production access, the degree to which activity can be attributed to a human or software principal, and the time required to revoke a session once it has been flagged.

Go deeper than the headline

Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.

Start your free 7-day trial

Sources

1 outlet

AlphaMaven summarised this story from the reporting below. Each link opens the original article.

TokenPost Shared PCs for people and AI call for access checks on every request