An IDE agent has employee permissions: what to do

Copilot now has managed models, a sandbox, and automatic SSO authentication. How to manage a coding agent like an employee with access rights.

  • What changed
  • The problem for companies: people choose models themselves, while access is granted by default
  • How it works technically
  • What this means for TTU

What changed

  1. Over three weeks, GitHub released four updates that, individually, look like changelog entries: managed default models in Copilot for JetBrains, a local sandbox for agents, automated SSO authorization for PATs and SSH keys, and the latest CodeQL.

  2. Together, they present engineering leaders with a new task: managing an agent like an employee with access rights. In Copilot for JetBrains, an enterprise administrator uses managed settings to choose the agent model that starts every new conversation.

  3. A developer can switch models manually in the picker, but the administrator sets the starting point. In the weekly release of September 21 Copilot now includes Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and Grok 4.7.

  4. Access depends on the plan: Opus 5.5 and GPT-6 Sol are available with Pro+, while Luna and Grok 4.7 are available with Pro.

  5. The same release introduced a local sandbox in the Copilot app.

  6. It limits the agent’s access to files, the network, and credentials.

  7. The third update concerns SSO authentication: an Enterprise Cloud administrator can enable delegation through GitHub Apps installed at the enterprise level.

  8. Then standard PATs and SSH keys are authorized in SSO organizations automatically, so developers do not need to do it manually in each organization.

The problem for companies: people choose models themselves, while access is granted by default

  1. When a company has a dozen models, every developer chooses their own.

  2. One chooses the most expensive, another the fastest, and a third the one they read about yesterday.

  3. Request cost, code quality, and the risk of context leakage vary from person to person, so the engineering director cannot forecast the budget. Access works the same way. The SSO changelog says it directly: manual authorization for each organization creates friction, so people create long-lived tokens to avoid reissuing them.

  4. An inconvenient process creates unsafe behavior.

  5. Now add an agent that reads files, accesses the network, and uses the same tokens. A long-lived PAT on a developer’s machine gains a second user: it works faster than a person and does not ask where it is allowed to go.

How it works technically

Three layers address three different risks. Model layer. Managed default sets the organization’s starting model. A powerful model is reserved for complex tasks, while routine conversations start with a cheaper one. The administrator makes this decision once, so developers do not have to repeat the choice in every session. Execution layer. The sandbox determines what the agent can access.

If it is restricted to specific directories, a list of network addresses, and a defined set of secrets, a prompt error or a substituted instruction in someone else’s README costs less.

This is the principle of least privilege we apply to service accounts, now extended to a tool on the developer’s laptop. Validation layer. CodeQL 2.27.2 remains a static analyzer that checks the result regardless of whether the code was written by a person or an agent. The Default set contains 498 security queries for

CWE Extended adds 131 more queries and

CWE. Version 2.27.2 adds std::regex parsing in C++, SQL injection models for the Comdb2 C API, and improvements for Go, Rust, and JavaScript. The more code an agent writes, the more important deterministic rule-based checks become instead of a second model’s opinion.

Assess where AI can deliver impact in your process

What this means for TTU

We measure the tool by time to use (TTU)

A model that produces a clean diff in one second is worthless if security spends the next two weeks determining where the agent went and which tokens it used.

Speed is measured until the change safely reaches production.

That is why governance reduces TTU, even though it looks like a restriction.

A centralized default model eliminates debates about the choice.

The sandbox eliminates the need to approve each agent individually.

Automatic SSO authorization frees developers from manual routine and the temptation to create a permanent token. Behind the image of “a developer opened the IDE and is working”

there are settings that someone must design and maintain.

What we do internally and in client projects

In AI-native development, we define the agent’s boundaries before the first line of code

In MCP integration projects, this means separate credentials for each server, short-lived tokens, and a call log.

Model requests go through the LLM & Security Gateway, which records the model, context size, and data that left the environment. In Python, PHP, .NET, and Node.js projects, static analysis in CI remains mandatory, and the agent’s code goes through it just like human-written code.

No new platform is needed for this.

You need someone to define the policy, encode it in the configuration, and verify that it works.

Verdict

In a year, companies will not differ by their model lineup, because everyone will have the same lineup. The winners will be those where the agent operates within known boundaries, at a known cost, with verifiable results. If your agent currently uses a token whose expiration date nobody remembers, you do not have an AI strategy. You have an employee without a job description.

Discuss the article: An AI Agent in the IDE Has Employee Access: What…

Enter your email or phone number so we can get back to you.

Send via: