You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit b894356
Browse filesBrowse the repository at this point in the historyBrowse files
docs: record the outcome of the upstream get-env report (corroborated, patched, not yet fixed)
The draft in this file was posted as modelcontextprotocol/servers#4882 and has been corroborated by two outside commenters and picked up by a third-party patch. Every fact in the new Outcome section was re-read from the GitHub API on 2026-09-29 rather than transcribed: the issue is open; @sattyamjjain adds the angle this draft did not have (get-env is annotated readOnlyHint: true, so hosts that auto-approve read-only tools run the full environment dump without asking) and points at three pre-existing PRs; @shleder confirms severity, names the credential classes, and proposes allowlisting plus name-pattern redaction; PR #4889 (Resolves #4882, 2 files, +209/-14, with its own unit tests and an AI-assistance disclosure) is open and unmerged.
The section is explicit about what must NOT be claimed: this was not the first report (issue #3986 predates it by five months and is still open), and nothing is fixed - upstream main still returns JSON.stringify(process.env, null, 2) and still declares readOnlyHint: true, both re-checked today. It also states the honest one-paragraph form of the chain, and what to do when a fixed version ships (re-run the probe, record before/after scorecards here). This is the evidence chain for the benchmark paper's availability section; the paper draft itself is not touched.
| patch |[PR #4889](https://github.com/modelcontextprotocol/servers/pull/4889), opened 2026-09-28 19:53:47Z, `Resolves #4882`, **open and not merged**|
108
+
| earlier report of the same weakness |[issue #3986](https://github.com/modelcontextprotocol/servers/issues/3986), opened 2026-04-19 by another user, still open |
109
+
110
+
**What the comments added, in their own terms.**
111
+
112
+
*@sattyamjjain: the issue duplicates #3986 and three PRs are already open for it (#4001 and #4164
113
+
restrict the output to a single key, #4009 adds a documentation warning), asking a maintainer to pick
114
+
one. Independently of the duplicate question, that comment adds an angle this draft did **not** have:
115
+
*"`get-env` is annotated with `readOnlyHint: true, destructiveHint: false`. Hosts that auto-approve
116
+
read-only tools will run the full env dump without asking."* — and suggests registering the tool only
117
+
behind an explicit opt-in, because the reference server is copied as a starting template.
118
+
*@shleder: confirms the severity and names the mechanism — environment variables routinely hold
119
+
`GITHUB_TOKEN`, `AWS_SECRET_ACCESS_KEY`, API keys and connection strings, so a prompt-injected or
120
+
unprompted agent call can exfiltrate credentials — and proposes exactly the two remediations that were
121
+
in the draft's third bullet: strict allowlisting, and redaction by name pattern (`KEY`, `TOKEN`,
122
+
`SECRET`, `PASSWORD`, `AUTH`). The comment discloses that its author maintains a sandbox layer product.
123
+
124
+
**What the patch does**, per its own description: `isSensitiveEnvVar` + `getRedactedEnv` in
125
+
`src/everything/tools/get-env.ts`, redacting to `[REDACTED]` on substring and prefix rules, plus expanded
126
+
unit tests; 2 files changed, +209/−14. It discloses AI assistance under human review.
127
+
128
+
**What is still true, checked rather than assumed:** upstream `main`'s
129
+
`src/everything/tools/get-env.ts` still returns `JSON.stringify(process.env, null, 2)` and still declares
130
+
`readOnlyHint: true`. So as of 2026-09-29 the exposure is unchanged in the code, the issue is open, and the
131
+
patch is open.
132
+
133
+
**How to describe this chain, in one honest paragraph.** The benchmark ran a fixed probe corpus against
134
+
five servers and flagged `get-env` on the Everything reference server as its highest-severity result; the
135
+
report was filed upstream as #4882 with a reproducible transcript, was independently corroborated by two
136
+
outside commenters — one adding the auto-approval (`readOnlyHint`) angle, the other specifying
137
+
allowlist/redaction — and a third-party patch (#4889) now resolves it, unmerged. It is **not** the first
138
+
report of this weakness (#3986 predates it by five months), and it is **not** fixed yet. What it is: an
139
+
independent reproduction with a harness that upstream contributors engaged with by name.
140
+
141
+
**Next, when a fixed version ships:** re-run `python -m bench run --target everything-public` and record
142
+
the before/after scorecards in this repository, and update this section rather than a chat log.
0 commit comments