AI agents aren't safe from prompt injection, and spreadsheets prove it
1–4 of 4 posts
Re: AI agents aren't safe from prompt injection, and spreadsheets prove it
#2is it really prompt injection to just muddle currencies on the page? I'd argue not - though it _is_ a clearly disingenuous quote.
Re: AI agents aren't safe from prompt injection, and spreadsheets prove it
#3is it really prompt injection to just muddle currencies on the page? I'd argue not - though it _is_ a clearly disingenuous quote.
It is an example of an indirect prompt injection: malicious prompt is injected in the content (spreadsheet file in this case) not in user prompt (instructions for the agent to pick the best offer). The file content, while looking benevolent to a human tricks an AI agent to perform actions (make decisions) in the interest of the attacker as opposed to the human user.
Other examples include job applications, refund requests, dispute remediation, etc. Virtually any situation in which a decision based on externally provided content is delegated to an AI agent.
This (indirect prompt injection) is a pretty serious threat. It is usually easy to pull off for attackers and yet it flies under the radar for users that are running such agents. In this case the agent can be properly sandboxed and use capable, modern models. The input files all look clean and reasonable. And yet the attacker can reliably trick the agent.
Re: AI agents aren't safe from prompt injection, and spreadsheets prove it
#4is it really prompt injection to just muddle currencies on the page? I'd argue not - though it _is_ a clearly disingenuous quote.
It is an example of an indirect prompt injection: malicious prompt is injected in the content (spreadsheet file in this case) not in user prompt (instructions for the agent to pick the best offer). The file content, while looking benevolent to a human tricks an AI agent to perform actions (make decisions) in the interest of the attacker as opposed to the human user. Other examples include job applications, refund req…
[flagged]