What happened
In January 2025, security firm Aim Security finds a flaw in Microsoft 365 Copilot, which it names "EchoLeak", and reports it to Microsoft. The flaw is fixed server-side and disclosed in June 2025 as CVE-2025-32711, rated critical. Microsoft says no customer action is required and no customer was affected.
The attack needs no click. The attacker sends the target an email written as if it were addressed to a human, never mentioning AI or Copilot: that is how it gets past the filters meant to spot planted instructions. Later, the employee asks Copilot a question on a topic the email mentions. Copilot fetches the relevant documents, this email included, and follows the instructions in it: gather the most sensitive information in its context and slip it into a link to the attacker's server, going through Microsoft addresses (Teams, SharePoint) that the protections let through.
Aim calls it an "LLM scope violation": untrusted data from outside drives access to internal data. The fix took five months.
Why it happened
For the model, a received email and a company instruction are the same text, in the same context. A rule such as "never obey instructions found in documents" assumes it can tell them apart; the attacker writes precisely so they look like an ordinary message.
And the assistant could, in the same turn, read outside content, read internal data and produce a link leaving the company. Those three together are enough for a leak.
The failure family
Two families of our catalogue: "Outside content" (obeying an order hidden in an email, a document or a tool response) and "The leak" (giving confidential information to the wrong person). The first opens the door, the second does the damage.
The code-level safeguard
- Outputs are filtered by the system: no link, no image, no sending to an address that is not on an approved list.
- Content from outside is marked as such: while it is in the context, tools that send data out are disabled or require confirmation.
- The assistant only sees the data the question needs, not everything the employee can read.
What we test
The "planted instruction", "hidden instruction" and "trap in the tool" tests slip an order into an email, a document or a tool response; the "data leak" test looks at whom the agent gives information to. The fault is a recorded action: data sent to an unverified address, whatever the agent says about it.
And your agent, what would it do?
The crash test puts your agent in these situations, dozens of times, on several models. You get the figures, the conversations where it gave in, and the fix plan.
Further reading
- Google Antigravity: the agent that wiped a drive to clear a cache
- Claude Code: the agent that thought it was doing a security audit
- Cursor: the AI support agent that invented a login rule
- Replit: the agent that wiped production during a code freeze
- Chevrolet: the chatbot that sold an SUV for one dollar
- Air Canada: held liable for a rule its chatbot made up