Research demo
OpenAI Agent Builder: a poisoned lead-form submission redirects a multi-agent sales workflow into emailing Salesforce CRM records to the attacker
AIR-2025-0149 AIR record · 2025-10-openai-agent-builder-mcp-exfil
From the supporting source
even if a model identifies a prompt injection—it can still be passed verbatim from one agent to another, allowing for the manipulation of agents deeper in the workflowpromptarmor.substack.com
The record
Disclosed
2025-10-14
Vendor
OpenAI
Product
OpenAI Agent Builder
Risk domain
Privacy › Data exfiltration channels
Also touches
Confused deputy
System type
Generative model
Who acted
Agent acted — the multi-agent workflow emailed CRM records to the attacker
Attack surface
Multi-agent system
Autonomy level
Semi-autonomous
Initial vector
Indirect prompt injection form submission
Impact type
Data exfiltration
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI02 – Tool Misuse and Exploitation
- MITRE ATLAS
- AML.T0024 – Exfiltration via AI Inference API
- NIST AI RMF
- MAP 4.2
- EU AI Act
- Article 10
- ISO/IEC
- TR 24028; 23894
- AIUC-1
- A. Data & Privacy
- Scope
- tool misuse
- Mitigation
- Monitor network traffic for anomalous patterns; Restrict egress destinations.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
Agent Builder workflows that ingest untrusted external submissions
Capability abused
Defense & remediation
Recorded defense
OpenAI guardrails and structured outputs
Guardrail kinds
Unknown
Guardrail outcome
Silent
Fix shipped
none; OpenAI acknowledged and recommended human-in-the-loop controls, noting protections 'aren't foolproof'
Taxonomy & references
Curator notes
PromptArmor. The multi-agent finding matters independently of the exfil: detecting an injection at the first hop does not stop it, because the flagged text is still relayed downstream. Same family as ForcedLeak (2025-09-salesforce-forcedleak) but against OpenAI's workflow builder.
Supporting sources
Cite as
Enkrypt AI. "OpenAI Agent Builder: a poisoned lead-form submission redirects a multi-agent sales workflow into emailing Salesforce CRM records to the attacker". AIR, AIR-2025-0149. https://agentincident.vercel.app/i/AIR-2025-0149/ (accessed 2026-10-19).