Vulnerability  ·  2026-10-22

Nous Research Hermes Agent webhook route-toolset privilege escalation via chat_id parsing (CVE-2026-114463)

VulnerabilityHigh impactGlobalCVE-2026-114463
Verified on the merged pull request (fetched): toolsets_for_source recovered the per-route toolsets grant by splitting chat_id on ':', allowing a route named 'build:external' to inherit 'build' toolsets; reproduced on main through the real aiohttp handler with a signed request. Reported in GHSA-2fmg-cjqm-hhrj.
Webhook-driven agent invocation is a standard integration pattern for agent gateways; this parse inconsistency lets an attacker holding one weak webhook credential reach another route's terminal/file tools — arbitrary command/file access on the agent host.
The webhook gateway builds session chat_id as 'webhook:{route_name}:{delivery_id}' and re-derives the authorizing route by splitting on ':', while authentication uses the exact URL segment; delivery_id is caller-supplied (X-GitHub-Delivery, svix-id, X-Request-ID), so a crafted route name like 'build:external' resolves to route 'build' and inherits its privileged toolsets.
Nous Research Hermes Agent — webhook platform gateway (fixed in PR #115285, GHSA-2fmg-cjqm-hhrj)
Apply the fix merged in PR #115285 (key toolsets_for_source on source.user_id stamped from the authenticated URL segment instead of splitting chat_id); update Hermes Agent.
Hermes Agent fix PR #115285 (fetched)NVD CVE-2026-114463
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →