GHSA-VG83-HCP4-5QCC
Vulnerability from github – Published: 2026-11-19 16:10 – Updated: 2026-11-19 16:10Summary
The domshell_eval MCP tool (and the eval command inside domshell_execute) was classified as read-tier, allowing it to run without the --allow-write server flag. However, eval and the write-tier js command dispatch to the same handler in the extension and call CDP Runtime.evaluate with the raw expression string and no side-effect gate. Any JavaScript expression could therefore mutate DOM, window, or global state through domshell_eval without the operator's --allow-write authorization.
The tool description explicitly promised "(read-only) ... always available in read-only mode" — the misclassification contradicted the documented contract, not just an implicit one.
Impact
No new capability was exposed — the same write access was already available via domshell_js with --allow-write — but the tier misclassification meant:
- MCP hosts (Claude Desktop, Cursor, etc.) rendered the wrong approval UI (
readOnlyHint: true) - Server-side tier flags (
--allow-write) misrepresented what an AI agent was actually authorized to do - Audit logs and deny messages labeled write-tier activity as read-tier
- Downstream tools and analysis relying on the tier separation could over-trust an
evalcall
Framing: OWASP LLM Top 10 LLM06 (Excessive Agency), NIST AI RMF (tool permission boundary + human-approval fidelity), CWE-863 (Incorrect Authorization).
The reporter's main framing: the MCP approval UI and server-side permission model can misrepresent the actual agency being granted to an AI agent.
Affected code paths
mcp-server/index.ts:113—WRITE_COMMANDSexcludedevalmcp-server/index.ts:879-887—domshell_evalregistered withANNO_READand description asserting read-onlysrc/background/index.ts:1849-1854—jsandevalboth dispatch tohandleJs(rawCode)src/background/index.ts:3282—handleJscallscdp.evaluateJs(code)src/background/cdp_client.ts:660-667—evaluateJssendsRuntime.evaluatewithexpression: code, returnByValue: true, awaitPromise: trueand nothrowOnSideEffectgate
Fix
Shipped in @apireno/domshell@2.0.8 (2026-07-08):
evaladded toWRITE_COMMANDS, aligning tier with actual capabilitydomshell_evalMCP annotation changed fromANNO_READtoANNO_WRITE- Tool description and
MCP_INSTRUCTIONSupdated to remove read-only language; read-only DOM inspection needs pointed atdomshell_cat/domshell_text/domshell_find - Server-only patch — no Chrome Web Store submission required; extension code unchanged
SECURITY.mdadded + GitHub Private Vulnerability Reporting enabled on the repo
Upgrade path
npm install @apireno/domshell@2.0.8
Or, in a persistent install, restart your MCP server against the new version.
Drives that call eval without --allow-write will start receiving Error: 'eval' requires --allow-write. Options:
- Add
--allow-writeto the server flags ifeval's write capability is actually needed - Migrate read-only inspection to
domshell_cat/domshell_text/domshell_find - Migrate to
domshell_jsif the caller already runs with--allow-write(same tier, same capability)
Integrator impact
- HKUDS/CLI-Anything — unaffected; their harness doesn't invoke
evalorjscommands (verified by grep). - Kgspin QA-UX drives — unaffected; they run with
--allow-writefor their action gates. - Direct Claude Desktop / Cursor sessions — any that called
domshell_evalwithout--allow-writewill now be denied, aligning behavior with the documented contract.
Reporter / Credit
Ugur Ozer, AI Risk Management https://airiskmanagement.ca info@airiskmanagement.ca
Source-review-first responsible disclosure via private email to the maintainer on 2026-07-08. Reporter finding id: DOMSHELL-EVAL-READ-TIER-BYPASS-001. Ugur followed the standard disclosure ladder correctly despite the missing SECURITY.md at time of report — that gap has now been closed (repo now has SECURITY.md + GitHub Private Vulnerability Reporting enabled).
Timeline
- 2026-07-08 — Report received by maintainer via private email
- 2026-07-08 — Source-side claims verified against
mainat2d63dab; fix planned - 2026-07-08 —
@apireno/domshell@2.0.8released to npm and the MCP registry;SECURITY.mdadded; GitHub PVR enabled - 2026-07-08 — This advisory published
{
"affected": [
{
"package": {
"ecosystem": "npm",
"name": "@apireno/domshell"
},
"ranges": [
{
"events": [
{
"introduced": "0"
},
{
"fixed": "2.0.8"
}
],
"type": "ECOSYSTEM"
}
]
}
],
"aliases": [
"CVE-2026-75518"
],
"database_specific": {
"cwe_ids": [
"CWE-266",
"CWE-863"
],
"github_reviewed": true,
"github_reviewed_at": "2026-11-19T16:10:53Z",
"nvd_published_at": "2026-11-18T22:17:24Z",
"severity": "MODERATE"
},
"details": "### Summary\n\nThe `domshell_eval` MCP tool (and the `eval` command inside `domshell_execute`) was classified as **read-tier**, allowing it to run without the `--allow-write` server flag. However, `eval` and the write-tier `js` command dispatch to the same handler in the extension and call CDP `Runtime.evaluate` with the raw expression string and no side-effect gate. Any JavaScript expression could therefore mutate DOM, window, or global state through `domshell_eval` without the operator\u0027s `--allow-write` authorization.\n\nThe tool description explicitly promised \"(read-only) ... always available in read-only mode\" \u2014 the misclassification contradicted the documented contract, not just an implicit one.\n\n### Impact\n\nNo new capability was exposed \u2014 the same write access was already available via `domshell_js` with `--allow-write` \u2014 but the tier misclassification meant:\n\n- MCP hosts (Claude Desktop, Cursor, etc.) rendered the wrong approval UI (`readOnlyHint: true`)\n- Server-side tier flags (`--allow-write`) misrepresented what an AI agent was actually authorized to do\n- Audit logs and deny messages labeled write-tier activity as read-tier\n- Downstream tools and analysis relying on the tier separation could over-trust an `eval` call\n\nFraming: OWASP LLM Top 10 LLM06 (Excessive Agency), NIST AI RMF (tool permission boundary + human-approval fidelity), CWE-863 (Incorrect Authorization).\n\nThe reporter\u0027s main framing: **the MCP approval UI and server-side permission model can misrepresent the actual agency being granted to an AI agent.**\n\n### Affected code paths\n\n- `mcp-server/index.ts:113` \u2014 `WRITE_COMMANDS` excluded `eval`\n- `mcp-server/index.ts:879-887` \u2014 `domshell_eval` registered with `ANNO_READ` and description asserting read-only\n- `src/background/index.ts:1849-1854` \u2014 `js` and `eval` both dispatch to `handleJs(rawCode)`\n- `src/background/index.ts:3282` \u2014 `handleJs` calls `cdp.evaluateJs(code)`\n- `src/background/cdp_client.ts:660-667` \u2014 `evaluateJs` sends `Runtime.evaluate` with `expression: code, returnByValue: true, awaitPromise: true` and no `throwOnSideEffect` gate\n\n### Fix\n\nShipped in `@apireno/domshell@2.0.8` (2026-07-08):\n\n- `eval` added to `WRITE_COMMANDS`, aligning tier with actual capability\n- `domshell_eval` MCP annotation changed from `ANNO_READ` to `ANNO_WRITE`\n- Tool description and `MCP_INSTRUCTIONS` updated to remove read-only language; read-only DOM inspection needs pointed at `domshell_cat` / `domshell_text` / `domshell_find`\n- Server-only patch \u2014 no Chrome Web Store submission required; extension code unchanged\n- `SECURITY.md` added + GitHub Private Vulnerability Reporting enabled on the repo\n\n### Upgrade path\n\n```\nnpm install @apireno/domshell@2.0.8\n```\n\nOr, in a persistent install, restart your MCP server against the new version.\n\nDrives that call `eval` without `--allow-write` will start receiving `Error: \u0027eval\u0027 requires --allow-write`. Options:\n\n- Add `--allow-write` to the server flags if `eval`\u0027s write capability is actually needed\n- Migrate read-only inspection to `domshell_cat` / `domshell_text` / `domshell_find`\n- Migrate to `domshell_js` if the caller already runs with `--allow-write` (same tier, same capability)\n\n### Integrator impact\n\n- **HKUDS/CLI-Anything** \u2014 unaffected; their harness doesn\u0027t invoke `eval` or `js` commands (verified by grep).\n- **Kgspin QA-UX drives** \u2014 unaffected; they run with `--allow-write` for their action gates.\n- **Direct Claude Desktop / Cursor sessions** \u2014 any that called `domshell_eval` without `--allow-write` will now be denied, aligning behavior with the documented contract.\n\n### Reporter / Credit\n\n**Ugur Ozer, AI Risk Management**\n\u003chttps://airiskmanagement.ca\u003e\n\u003cinfo@airiskmanagement.ca\u003e\n\nSource-review-first responsible disclosure via private email to the maintainer on 2026-07-08. Reporter finding id: `DOMSHELL-EVAL-READ-TIER-BYPASS-001`. Ugur followed the standard disclosure ladder correctly despite the missing SECURITY.md at time of report \u2014 that gap has now been closed (repo now has SECURITY.md + GitHub Private Vulnerability Reporting enabled).\n\n### Timeline\n\n- 2026-07-08 \u2014 Report received by maintainer via private email\n- 2026-07-08 \u2014 Source-side claims verified against `main` at `2d63dab`; fix planned\n- 2026-07-08 \u2014 `@apireno/domshell@2.0.8` released to npm and the MCP registry; `SECURITY.md` added; GitHub PVR enabled\n- 2026-07-08 \u2014 This advisory published",
"id": "GHSA-vg83-hcp4-5qcc",
"modified": "2026-11-19T16:10:53Z",
"published": "2026-11-19T16:10:53Z",
"references": [
{
"type": "WEB",
"url": "https://github.com/apireno/DOMShell/security/advisories/GHSA-vg83-hcp4-5qcc"
},
{
"type": "ADVISORY",
"url": "https://nvd.nist.gov/vuln/detail/CVE-2026-75518"
},
{
"type": "WEB",
"url": "https://github.com/apireno/DOMShell/commit/03c1d7625e220297aff0dde37613d65cfe6f9a36"
},
{
"type": "PACKAGE",
"url": "https://github.com/apireno/DOMShell"
},
{
"type": "WEB",
"url": "https://github.com/apireno/DOMShell/releases/tag/mcp-v2.0.8"
}
],
"schema_version": "1.4.0",
"severity": [
{
"score": "CVSS:4.0/AV:L/AC:L/AT:N/PR:L/UI:N/VC:L/VI:L/VA:N/SC:N/SI:N/SA:N",
"type": "CVSS_V4"
}
],
"summary": "DOMShell: domshell_eval classified as read-tier can execute side-effecting JavaScript without --allow-write"
}
Sightings
| Author | Source | Type | Date | Other |
|---|
Nomenclature
- Seen: The vulnerability was mentioned, discussed, or observed by the user.
- Confirmed: The vulnerability has been validated from an analyst's perspective.
- Published Proof of Concept: A public proof of concept is available for this vulnerability.
- Exploited: The vulnerability was observed as exploited by the user who reported the sighting.
- Patched: The vulnerability was observed as successfully patched by the user who reported the sighting.
- Not exploited: The vulnerability was not observed as exploited by the user who reported the sighting.
- Not confirmed: The user expressed doubt about the validity of the vulnerability.
- Not patched: The vulnerability was not observed as successfully patched by the user who reported the sighting.
The approach is described in our paper Mapping CVEs to MITRE ATT&CK Techniques: A Curated Gold-Set Classifier and the Limits of LLM-Assisted Label Expansion.
Browse all ATT&CK techniques and the vulnerabilities related to each.
Related by attack behaviour
Vulnerabilities whose description is nearest to this one in the vector space of the CIRCL/vulnerability-attack-technique-biencoder model. This is a similarity search over the bi-encoder space (plain cosine), not a classification, and it has no measured accuracy.