One Email Could Turn Your AI Assistant Into an Insider Threat
An attacker can hide instructions in material an AI assistant is asked to read. Two documented flaws show how that can put company data at risk, and where the security boundary belongs.

Imagine asking an AI assistant to catch you up on today’s email. It opens a routine-looking message about an HR task. Buried in that message is an instruction to find an employee’s details elsewhere in your inbox and send them to a supposed “compliance” website. You never asked the assistant to do that. The sender of the email did.
This is indirect prompt injection. The attacker does not control your request to the assistant. They control lower-trust material the assistant encounters while carrying out that request: an email, web page, shared document, or tool result. The attack works if the assistant treats that material as an instruction rather than information to examine. Microsoft’s security researchers describe data theft and unauthorized actions as possible outcomes.
A real flaw showed how far it could go
In 2025, researchers disclosed EchoLeak, a vulnerability in Microsoft 365 Copilot. Their crafted email could, under certain conditions, cause Copilot to expose limited internal data that the user was already permitted to access. The victim did not have to click a malicious link for the technique to work. Microsoft says it has fixed the vulnerability, tracked as CVE-2025-32711. EchoLeak was a research demonstration, not evidence that every Copilot deployment was breached. Microsoft’s account and the research paper document the case.
Another 2025 demonstration, ShadowLeak, targeted ChatGPT’s Deep Research agent when connected to Gmail. Researchers placed instructions in an email that appeared to describe a normal HR workflow. When the agent processed the inbox, the researchers showed it could retrieve personal details from another message and send them to an attacker-controlled destination through its browsing tool. The researchers reported that the flaw was resolved; OpenAI later discussed the example as part of its work on agent security. Radware’s original report describes the test and disclosure.
These examples expose the same boundary: the assistant may have legitimate access to private information, while the attacker has access only to something the assistant will read. If the attacker’s text can steer the assistant across that boundary, ordinary access permissions alone do not settle whether the action is authorized.
Why “ignore malicious instructions” is insufficient
A hidden command does not have to say “ignore previous instructions.” It can sound like a normal part of the job: verify a record, complete a report, open a reference link. OpenAI says effective attempts increasingly resemble social engineering, with false claims of authority and plausible business context. That is why simply screening for suspicious phrases cannot reliably solve the problem.
The risk grows when one assistant can both read sensitive data and take actions—visit outside websites, send messages, or call connected tools. OWASP lists prompt injection as its top large-language-model application risk and recommends limiting privileges, validating actions in application code, and requiring human approval for high-risk operations. OWASP’s guidance also cautions that retrieval techniques and model tuning do not eliminate the vulnerability.
For a business deploying an AI assistant, the practical questions are concrete: Which files and messages can it read? Which external destinations can it contact? Can it send information without a person reviewing the recipient and contents? Are those decisions enforced by the surrounding software, or merely requested in the assistant’s instructions?
The documented flaws above were fixed. The underlying design challenge remains. An assistant that reads untrusted material while holding broad access needs controls at the point where it uses that access. Otherwise, the next person giving it instructions may be the author of the email it was supposed to summarize.