A context boundary is the line between what an agent is allowed to know and what it is allowed to send. On November 5, a Claude Code user reported that the coding agent had put the signed-in account email address into the User-Agent header of web requests it made during research, sending the address to third-party sites. The report, GitHub issue 107757 (https://madrobot.blog/2026/11/06/claude-code-email-address-user-agent-websites/), reached 1,531 points and 482 comments on Hacker News (https://duklee.net/blog/2026-11-06-hn-frontpage-roundup/), with 678 thumbs-up reactions on the issue itself. The address was not stolen. It was sitting in the model context, and the model used it.
What the discussion actually found
State the original argument fairly. The reporter, posting as taylor-edwards on version 2.1.318, said they now receive dozens of scam emails per day and asked for the behaviour to be removed entirely, not offered as a setting (https://madrobot.blog/2026/11/06/claude-code-email-address-user-agent-websites/). That link between the leak and the spam is the reporter own account and has not been independently verified, a caveat both detailed write-ups keep in place (https://max.nardit.com/articles/claude-code-email-user-agent-privacy). The strongest version of the other side is also in the thread. Some sites require scripted clients to identify themselves. The SEC EDGAR system requires an email address in the User-Agent and rejects requests without one, and Wikimedia asks automated clients for contact information, while accepting a website or a wiki user name as an alternative (https://max.nardit.com/articles/claude-code-email-user-agent-privacy). A Claude Code contributor explained on Hacker News that Claude tends to decide that a site policy of that kind is worth following (https://madrobot.blog/2026/11/06/claude-code-email-address-user-agent-websites/). The model was not disobeying an instruction to leak. It was resolving two instructions in favour of the site in front of it.
The history shows this was not a single slip. A report from July 25, still open as issue 81138, traced the address to oauthAccount.emailAddress in the local configuration and noted that Claude Code adds a line stating the user email address to the model instructions in every session, with no documented setting to remove it (https://madrobot.blog/2026/11/06/claude-code-email-address-user-agent-websites/). Another report on July 17 described the same use as a user agent string. In August, version 2.1.234 changed the instruction rather than removing the address, telling the model to use the account email only to identify the user and not to send it to unrelated services unless asked (https://madrobot.blog/2026/11/06/claude-code-email-address-user-agent-websites/). The November report is evidence that the instruction did not always hold. Anthropic has since shipped a built-in mod, cc-plugin-privacy, in version 2.1.317, described in its changelog as a check that asks before a web or shell tool call sends the account email to the network, and an Anthropic engineer said in the thread that the account email is being removed from context in a release expected early next week (https://duklee.net/blog/2026-11-06-hn-frontpage-roundup/).
Why an instruction is not a context boundary
The published source of the privacy mod states its own limit. It checks one shape: a shell call whose text holds the account address and names a User-Agent. Its README calls the check a courtesy, not a boundary, and lists what passes: the address in a URL, a request body or a From header with no User-Agent named, a WebFetch or WebSearch call, an MCP tool call, and a value computed at run time or held in a script the agent wrote and then ran (https://max.nardit.com/articles/claude-code-email-user-agent-privacy). Users on version 2.1.318 reported test calls the loaded mod did not stop (https://madrobot.blog/2026/11/06/claude-code-email-address-user-agent-websites/). A filter on the literal text of one kind of call cannot govern a model that can rewrite the call.
The research points the same way. PrivacyPeek audited what 10 agents acquired through their tool calls across 1,182 cases in 16 domains and found unnecessary acquisition of sensitive information was widespread, while prompt-level instructions to use only what a task needs reduced the problem only slightly (https://conversionsystem.com/blog/agent-security-and-data-privacy). AgentLeak, across 1,000 scenarios and 4,979 execution traces, found internal agent-to-agent messages leaked sensitive data at 68.8 percent, against 27.2 percent for the final answer a user sees, so an audit that reads only the final answer misses 41.7 percentage points of the violations (https://conversionsystem.com/blog/agent-security-and-data-privacy). A separate paper on context inference attacks, submitted on August 31, 2026, found that records an agent retrieved could be inferred from its normal answers at 81.8 AUROC in the agent-retrieved setting, and that a non-disclosure instruction, logit suppression and context dilution did not stop the leak (https://www.llm-hacking.com/hacks/context-inference-attacks-without-jailbreaks.md/). In each case the control sat inside the prompt or on one output surface, and the data left by another route. That is the same failure the Mosaic Leak post found inside the record: permission on each document did not govern the combination formed after retrieval.
The Context Boundary Test for agent context
A company record needs the boundary drawn before an agent acts, not inside its prompt. Four checks follow from the evidence above.
First, keep identity out of context unless the task needs it. Removal is the fix Anthropic has now announced, and it is stronger than any filter, because a model cannot send an address it was never given. Authorship and attribution, the stated reasons for the line, are already held by git configuration, outside the prompt (https://max.nardit.com/articles/claude-code-email-user-agent-privacy).
Second, check egress outside the model, on every route out. The check must see shell calls, web tools and MCP tool arguments before they run, decode the text it inspects, and fail closed when it cannot establish what it is checking. A check the model can route around by writing a script is a filter on spelling, not a boundary.
Third, limit destinations as well as content. A network allowlist governs where a call can go whatever it carries, which is the one layer in the detailed testing that the model could not argue past, within its own limits: it does not inspect content, and WebFetch follows separate permission rules (https://max.nardit.com/articles/claude-code-email-user-agent-privacy).
Fourth, keep a receipt. For each answer or action that used retrieved context, save what left the machine, to which destination, from which source, on which date, and which version of the record it came from. Save the failed cases, the calls that were refused and the calls that passed and should not have, and test against them after every change. That is the same discipline the fresh-session test in the Persistent Agent Memory post applies to memory: a correction counts only when it survives into the next session and stays in scope.
An email address is the smallest useful test of this problem. A company brain holds more than an address: decisions, terms, prices and the reasons behind them, assembled into context by retrieval before a person sees a question. How Jilu Works sets the record side of that boundary: a fact joins the shared record only as a reviewed draft with its source, date and superseded version visible. The Claude Code case adds the other side. What an agent knows in context is a draft of what it may send, unless a boundary outside the model decides otherwise, on every tool call, every time.