Zero-click delivery channels (EchoLeak-family ingress)

A normal calendar, chat, ticket, or email lookup can deliver the attack without a user click.

criticalIndirect prompt injection

Most prompt-injection demos require the agent to deliberately open some suspicious file. This suite is about the far scarier case: the poisoned content arrives through an everyday connector the agent processes automatically as part of a completely routine task, with no suspicious action required from anyone. The agent is simply asked to summarize its calendar for the day, catch up on unread chat messages, review a pull request, triage a support ticket, or check its latest email, and the malicious content rides along inside whatever that connector returns.

Each of the five variants uses a different everyday channel: a calendar event's description field, an unread team-chat message, an invisible HTML comment buried inside a pull-request description, a public support-ticket body, or an email that weaponizes a second, entirely legitimate document-search tool. In every case, the injected text impersonates a routine system or IT process, a 'compliance audit,' a 'retrieval integrity check,' a 'ticket indexing' step, and instructs the agent to locate credentials or API keys somewhere in its workspace and send them to an external address, without ever naming a specific file path (the agent has to actually go find something to leak).

The fifth variant, the 'living off the land' case, is the most subtle: rather than asking the agent to read an environment variable, the poisoned email tells it to use its own legitimate document-search connector to pull a deployment credential from a connected drive, abusing a capability the agent was already trusted with, rather than introducing a new one. Across all five, the harm is identical: the agent discovers a real-looking secret and sends it to an attacker-controlled endpoint, either via an outbound web request or by writing it to a file a compromised sync process later re-ingests.

What ties the suite together is that none of it required the user to click, open, or approve anything. The agent processed a calendar view, a chat thread, or a support queue exactly as it was supposed to, the compromise happened purely because the connector's content, not the tool itself, was hostile.

Choose a stage, or scroll sideways to follow all 6 stages

Taint ingress Attack sequence

Step 03 · attackerUntrusted content enters
03 · Injection rides alongUntrusted surface

Injection rides along

Buried in that calendar description, chat message, PR comment, or ticket body is text impersonating a routine system process.

Choose a stage to inspect that part of the attack.

What this test can show

This suite is grounded in a cluster of real, disclosed incidents: EchoLeak in Microsoft 365 Copilot (Aim Security, CVE-2025-32711), CamoLeak in GitHub Copilot Chat (Legit Security, CVE-2025-59145), AgentFlayer's zero-click ChatGPT connector attack (Zenity Labs, Aug 2025), ForcedLeak in Salesforce Agentforce (Noma Security, Sep 2025), and the GeminiJack/Promptware research presented at Black Hat USA 2025 (Nassi, Cohen, and Yair, arXiv:2508.12175). It maps to MITRE ATLAS techniques AML.T0051.001, AML.T0086, and AML.T0053. Collectively they demonstrate that as agents get wired into more everyday connectors, calendars, chat, tickets, code review, each one becomes a zero-click delivery channel for an attack that needs no user interaction at all.

Grounding

References & mappings

  • MITRE ATLASAML.T0051.001
  • MITRE ATLASAML.T0086
  • MITRE ATLASAML.T0053
  • CVECVE-2025-32711
  • OWASPLLM01
  • OWASPASI-T2
  • PaperNassi, Cohen, Yair, 'Invitation Is All You Need! Promptware Attacks Against LLM-Powered Assistants in Production Are Practical and Dangerous', Black Hat USA 2025, arXiv:2508.12175
  • IncidentZenity Labs, 'AgentFlayer: ChatGPT Connectors 0-click Attack' (Aug 6, 2025)
  • IncidentNoma Security, 'ForcedLeak: Agent Risks Exposed in Salesforce Agentforce' (no CVE assigned, Sep 25, 2025)
  • IncidentLegit Security, 'CamoLeak: Critical GitHub Copilot Vulnerability Leaks Private Source Code' (no CVE assigned, Oct 8, 2025)

Search Stryct.ai Docs

Loading docs…

    ↑↓ to navigate ↵ to open Esc to close