Targets the 2026 enterprise automation attack surface: n8n · Zapier · Make.com · Power Automate · Salesforce Agentforce · Microsoft Copilot Studio · ServiceNow Now Assist. These platforms run with high privilege, direct access to CRM data, email, SharePoint, ITSM systems — the perfect pivot point. SPECTER RELAY fingerprints, exploits, hijacks OAuth tokens, builds rogue implants that survive deactivation, and exfiltrates data via the platform's own trusted connectors.
Every CVE is implemented as a working PoC payload builder, not a scanner. INJECT gate required for live fire.
| CVE / TTP | Platform | CVSS | Technique |
|---|---|---|---|
| CVE-2026-21858 "Ni8mare" | n8n | 10.0 | Unauth RCE via Content-Type confusion, webhook form endpoint |
| CVE-2025-68668 "N8scape" | n8n | 9.9 | Pyodide sandbox escape via ctypes — runs in n8n worker process |
| CVE-2025-68613 | n8n | 9.8 | Expression injection RCE via jmespath sink |
| CVE-2025-68697 | n8n | — | Legacy JS code node arbitrary file read/write |
| CVE-2026-21877 | n8n | 9.9 | Workflow RCE chain |
| CVE-2025-32711 "EchoLeak" | Copilot Studio / M365 | 9.3 | 0-click RAG injection via inbound email — silently exfiltrates mailbox |
| CVE-2026-21520 "ShareLeak" | Copilot Studio | 7.5 | SharePoint connector exfil via adaptive card form injection |
| CVE-2025-12420 | ServiceNow | 9.3 | Second-order agent-to-agent injection — bypasses MFA/SSO |
| CVE-2026-40374 | Power Automate Desktop | — | Secrets disclosed in %TEMP% |
| ForcedLeak (CVSS 9.4) | Agentforce | 9.4 | Web-to-Lead CSP whitelist expired domain hijack → CRM exfil |
| CoPhish | Copilot Studio | — | OAuth token theft via Sign-in topic — Datadog Labs 2025 |
| Prompt Mines | Agentforce | — | 0-click data corruption via hidden zero-width char instructions in records |
| Zenity DLP Bypass Quartet | Power Platform | — | 4 techniques bypassing Microsoft Power Platform DLP policies |
| AgentFlayer | Cross-platform | — | Zero-click cross-platform chain — Zenity Black Hat 2025 |
| Agent Session Smuggling | A2A | — | Stateful A2A injection between orchestrator/tool — Unit42 PAN 2026 |
Each subsystem is independently addressable from the CLI. SURVEY and HARVEST are OPEN-gate. INJECT, HIJACK, CHAIN require INJECT gate. PERSIST and live EXFIL require UNLEASHED.
Multi-platform fingerprinting across 7 enterprise automation targets. HTTP header/path probing, version detection, CVE applicability mapping (by n8n version). ServiceNow agent discovery. Power Automate Desktop secrets path enumeration.
Parses exported workflow JSON, Power Platform solution.zip, Salesforce metadata XML. 14 secret patterns — OpenAI, Anthropic, AWS, GitHub, Slack, Stripe, SendGrid, Twilio, Zapier MCP mcp.json. Connected service enumeration.
Working PoC payload builders for all 15 CVEs/TTPs listed above. InjectPayload dataclass with full headers, method, path, gate. Async fire() for live execution (INJECT gate). Power Platform DLP bypass quartet returns 4 separate payloads.
CoPhish — Copilot Studio Sign-in topic with attacker redirect URI. Multi-tenant Entra app manifest + phishing consent URL. Token validation via Microsoft Graph. Zapier MCP mcp.json credential dump via prompt injection.
NetworkX DiGraph of agent trust relationships. Dijkstra escalation path finder by privilege delta. ServiceNow CVE-2025-12420 second-order pivot. AgentFlayer cross-platform chain builder. Agent Session Smuggling (Unit42). JSON-LD export for court-admissible reporting.
n8n cron rearm — cron node self-reactivates after manual deactivation. Power Automate hidden dev environment clone (weaker DLP). Copilot Studio cross-tenant bot share (survives victim deletion). Make.com restartOnError=true. Agentforce SYSTEM_CONTEXT Apex Connected App. ServiceNow server-side business rule. All UNLEASHED-gated.
CSP whitelist hunter (ForcedLeak generalisation) — scans platform allowlists for expired/registerable domains via DNS + HTTP. ConnectorExfil: Agentforce CRM exfil (ForcedLeak), Copilot SharePoint (ShareLeak), Power Automate unblockable connector, n8n HTTP node, ServiceNow gs.sendEmail. Canary staging at INJECT level.
RLY-{hex12} report IDs. Ed25519 signing via PyNaCl. JSON-LD output with @context, finding counts, WMD class counts, evidence hash (SHA-256). Verify() checks signature integrity. Full finding metadata: MITRE ATLAS, OWASP, CVE, CVSS, remediation.
AML.T0051 — LLM Prompt Injection
AML.T0054 — LLM Jailbreak
AML.T0020 — Poison Training Data
AML.T0043 — Craft Adversarial Data
AML.T0048 — Exfiltration via ML Inference API
AML.T0049 — Evade ML Model
LLM01 — Prompt Injection
LLM02 — Sensitive Information Disclosure
LLM06 — Excessive Agency
LLM08 — Excessive Permissions
Agentic: AST01 / AST02 / AST03 / AST05 / AST07 / AST08 / AST09
MCP Top 10 2026: Tool Poisoning, Credential Exposure