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.
No new capability was exposed — the same write access was already available via domshell_js with --allow-write — but the tier misclassification meant:
readOnlyHint: true)--allow-write) misrepresented what an AI agent was actually authorized to doeval callFraming: 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.
mcp-server/index.ts:113 — WRITE_COMMANDS excluded evalmcp-server/index.ts:879-887 — domshell_eval registered with ANNO_READ and description asserting read-onlysrc/background/index.ts:1849-1854 — js and eval both dispatch to handleJs(rawCode)src/background/index.ts:3282 — handleJs calls cdp.evaluateJs(code)src/background/cdp_client.ts:660-667 — evaluateJs sends...2.0.8Exploitability
AV:LAC:LAT:NPR:LUI:NVulnerable System
VC:LVI:LVA:NSubsequent System
SC:NSI:NSA:N4.8/CVSS:4.0/AV:L/AC:L/AT:N/PR:L/UI:N/VC:L/VI:L/VA:N/SC:N/SI:N/SA:NAccess Control
Other