Disclosed vulnerability
n8n September 30 advisory batch: cross-user hijack of pending agent tool approvals, agent-node credential check bypass, MCP validation takeover and cross-project agent transfer
AIR-2026-0267 AIR record · 2026-09-n8n-agent-approval-hijack-advisory-batch
From the supporting source
any member of that project could list another member's private agent conversations, read them in full, and then answer a pending approval on their behalfgithub.com
The record
Disclosed
2026-09-30
Vendor
n8n
Product
n8n (Agents module, Message an Agent node, instance MCP server)
Risk domain
Access control and permissions › Privilege escalation
Also touches
Credential theft
System type
Generative model
Who acted
AI was the target (contested) — authorization flaws in an automation platform, several in the agent tool-approval path; contested because the agent's approval gate is the target
Attack surface
Agent framework
Autonomy level
Copilot with approval
Initial vector
Authenticated low privilege member abuse
Impact type
Unauthorized tool execution, Credential exposure, Account takeover, Unauthorized data access
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI03 – Identity and Privilege Abuse
- MITRE ATLAS
- AML.T0055 – Unsecured Credentials
- NIST AI RMF
- GOVERN 6.1
- EU AI Act
- Article 13
- ISO/IEC
- TR 24028; 42001; 23894
- AIUC-1
- B. Security
- Scope
- tool misuse
- Mitigation
- Apply least-privilege principle; Regular policy audits.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
n8n instances before 2.41.4 / 2.42.1 (1.123.83 on the 1.x line per a third-party summary) with shared projects, the Agents module or the instance MCP server enabled
Capability abused
Defense & remediation
Recorded defense
human tool-approval gate; shared-workflow credential check; MCP permission scopes
Guardrail kinds
Human approval, Permission scope
Guardrail outcome
Silent
Fix shipped
n8n 2.41.4 and 2.42.1 (2.42.0 fixes only the approval-hijack issue)
Taxonomy & references
Curator notes
One record for the coordinated 30 September 2026 batch, which a third-party summary counts as 14 advisories (10 high, 4 medium); no CVE identifiers had been assigned in the advisories reviewed. The record is built around four agent-specific advisories. GHSA-p3pg-xw4f-m72c (medium): the endpoint that resumes a suspended agent tool call did not check that the checkpoint belonged to the caller, so a project member could answer another member's pending approval and the tool ran. This is why autonomy is coded copilot_with_approval. GHSA-x25p-9mr6-cwgp (high): the shared-workflow credential check did not inspect the agent node's parameter-held model and tool credential ids, so an editor could have the owner's credentials resolved for tools they control and send decrypted secrets to a destination of their choosing. GHSA-5jr4-xmvf-frmj (high): the MCP workflow-validation tool evaluates caller-supplied code through an interpreter that allows assignment onto shared prototypes; a member with only a read grant could drive this to takeover of the instance owner's account (MFA-protected accounts are not affected). GHSA-2r5r-xgvc-rj4p (medium): the Agents create endpoint accepted a caller-supplied identifier, letting a read-only user move another project's agent into their own project. The remaining advisories in the batch are not individually described here. The advisories describe vulnerabilities and fixes; no exploitation is reported. Separate from earlier n8n records, which cover different mechanisms.
Supporting sources
- https://github.com/n8n-io/n8n/security/advisories/GHSA-p3pg-xw4f-m72c
- https://github.com/n8n-io/n8n/security/advisories/GHSA-x25p-9mr6-cwgp
- https://github.com/n8n-io/n8n/security/advisories/GHSA-5jr4-xmvf-frmj
- https://github.com/n8n-io/n8n/security/advisories/GHSA-2r5r-xgvc-rj4p
- https://wz-it.com/en/blog/n8n-security-update-september-2026/
Cite as
Enkrypt AI. "n8n September 30 advisory batch: cross-user hijack of pending agent tool approvals, agent-node credential check bypass, MCP validation takeover and cross-project agent transfer". AIR, AIR-2026-0267. https://agentincident.vercel.app/i/AIR-2026-0267/ (accessed 2026-10-19).