Driven
Back to resources

Security Perspective

Your AI Agent Knows Too Much

A reported email disclosure shows why enterprise AI needs controls over what leaves the business, not just what appears in chat.

THE CONTEXT

A small header. A larger warning.

A useful agent can complete the assignment and still disclose information it was never authorized to share. The enterprise question is not only whether the answer is correct. It is whether every action respects the boundary around your data.

On November 5, 2026, a user filed issue #107757 in Anthropic's Claude Code repository, reporting that the agent had included their email address in outbound HTTP User-Agent headers without permission. The reporter also attributed an increase in scam email to the disclosure. That causal link is an allegation, not an independently established finding.

A repository contributor described the behavior as a model-generated choice rather than an intentionally implemented email-sharing feature: an agent researching a website could supply the user's address when constructing a request that identified its operator. The contributor said a built-in privacy check was enabled by default in version 2.1.317 and later. Subsequent commenters reported gaps. This is not evidence that every request exposes an email, nor proof of a universal fix.

The distinction matters. The issue concerns information in a network request, not necessarily the response the employee sees. Public material reviewed for this article does not establish the number of affected users, every destination involved, or a confirmed compromise of Anthropic's infrastructure. The reported behavior is serious enough to investigate without inflating it into a different incident.

A contact requirement is not consent

Wikimedia's published policy asks automated clients to identify themselves and provide contact information. It accepts options including an email address, website, or wiki user. That does not require an employee's personal account address. Organizations should supply an approved service identity when needed, rather than let an agent decide which private identity to disclose.

The decision to share an identity belongs to the business, not to the model completing a task.

THE ANALYSIS

Helpful is not the same as authorized

Consider an illustrative enterprise task: a developer asks an assistant to research an integration. The assistant has useful project context, access to local files, and permission to fetch documentation. A remote service asks automated clients for identification. To finish the task, the assistant assembles a request using information already available to it. The research succeeds. An unnecessary disclosure may happen along the way.

This is a failure of boundaries, not simply a poor answer. Permission to use information for one purpose does not automatically permit sending it to another organization. Account identity, internal project names, customer references, and task descriptions should not become convenient material for constructing outbound requests. Availability inside a session is not authorization to publish.

The scenario does not need a malicious website to be risky. A legitimate operational requirement can conflict with a company's privacy expectations. Prompt injection adds another possible route, but assuming every failure requires an attacker misses the more routine problem: an agent can make a plausible decision that nobody approved.

The chat window is only one surface

A review limited to prompts and final answers will not tell you what a generated script, command, connector, or browser request transmitted. A model gateway can govern traffic that passes through it; it cannot, by itself, inspect a separate tool's direct connection to a website. Map both paths before declaring the workflow protected.

Our assessment is that agent adoption changes the approval unit. Teams must govern the proposed action, destination, and data together. The right question is not simply whether an agent may browse. It is which service it may contact, for what business purpose, and with which information.

Review the complete action: the destination, the information being sent, and the authority to send it.

WHAT MATTERS

Put the boundary outside the model

Start by removing unnecessary personal and confidential information from agent context. Where a workflow needs authentication, keep credentials in a controlled execution layer and expose only the minimum capability. Separate an employee's identity used for accountability from the service identity used for an approved external request.

Replace open-ended network access with a documented destination policy. For routine integrations, prefer structured tools with defined inputs over unrestricted command generation. Where commands are necessary, use an isolated runtime and enforce network restrictions independently of the model. OWASP identifies excessive functionality, permissions, and autonomy as distinct sources of agent risk; reducing one does not automatically resolve the others.

Validate what the selected control actually covers. Anthropic's sandbox documentation distinguishes shell-process isolation from tools, MCP servers, and hooks that run outside that boundary. A sandbox label alone is not proof of complete coverage. Check each execution path, including generated scripts and connector processes, against the intended policy.

Inspect the request, not just the command

For sensitive workflows, enforce approved request fields and minimize outgoing content in a controlled tool or properly configured inspection layer. A hostname allowlist limits destinations; it does not decide whether sending a customer identifier to an allowed host is appropriate. Command-text checks can add protection, but should not be treated as a complete view of what a program will send.

01

CONTEXT:

Give the agent only the data needed for its current task.

02

EXECUTION:

Constrain the tools, permissions, and destinations it can use.

03

DISCLOSURE:

Check the information an approved action will actually transmit.

OPERATIONAL RESPONSE

Make safe execution demonstrable

Begin with one representative workflow, not a company-wide promise. In a controlled test environment, give it synthetic identity and project data, then exercise ordinary research and connector actions. Inspect the received requests at a test endpoint. Confirm that prohibited fields are absent and that blocked actions produce a clear explanation rather than a silent workaround.

Repeat the test after model, agent, plugin, or policy updates. Record the deployed versions, allowed destinations, observed requests, and expected decisions. Measure both sides: unauthorized disclosure must fail, and approved work must remain usable. A control that routinely interrupts harmless tasks invites broad exceptions that undo its value.

Own the operating response

Assign an owner for the workflow and its exceptions. Retain enough redacted evidence to connect a request to its initiating user, tool, destination, and policy decision without turning monitoring into another store of sensitive data. If exposure is suspected, preserve evidence, establish what left and where it went, and involve security and privacy teams. An exposed email and an exposed credential require different responses.

Driven can help assess agent workflows, identify unnecessary data access, design execution and outbound-data controls, and validate them through practical testing. The objective is not to stop useful automation. It is to make the limits on that automation explicit, enforceable, and reviewable.

Reach out to Driven to review what your AI agents can access, what they can send, and how to enforce the boundaries your business needs.