Answers
Can prompt injection make your AI agent post to X?
Prompt injection can make an AI agent write a social media post you never asked for, and an agent that holds the power to publish will post it. Evenpost, an app for X (Twitter) that lets an AI agent draft posts you approve on your phone, gives the agent no post tool, so the worst an injected instruction can do is leave a draft for you to reject.
Evenpost / Answers / Prompt injection and posting
Prompt injection is text that reads like an instruction to your agent, hidden in something it was only meant to read: a web page it summarises, a README, an issue comment, a file. An agent with access to X can be told, by that text, to post a link, praise a product or announce something false, and it may not mention that it did.
What prompt injection can do to an agent that posts
- With a post toolThe injected text becomes a live post the moment the agent acts on it. A rule in the prompt, such as "never post without asking", is just more text for the injection to argue with.
- Through EvenpostThe agent has no approve tool and no post tool, and an Evenpost key cannot reach one. The injected text becomes a draft in your queue, nothing more.
- On your phoneYou read the exact words X would get, links included, with hidden characters such as zero-width spaces and direction marks removed. A draft you did not expect gets one tap: Reject.
- After you editOnce you edit a draft, your version is kept. An agent write that would change it is refused, so text the agent reads later cannot rewrite what you approved.
What else limits the damage
If the agent changes a draft while you are looking, the card says "Updated by your agent" and you read it again before it can go. Notifications carry fixed wording, never the draft's text, so injected words never reach your lock screen. A leaked Evenpost key can fill your drafts, never your timeline, and you can revoke it in Settings.
What is prompt injection?
Prompt injection is text written to be read by an AI agent as an instruction, hidden inside content the agent was only meant to read, such as a web page, a document or a comment. The agent cannot always tell your request from the planted one, which is why what it is able to do matters most.
Can a rule in the prompt stop an agent from posting?
A rule in the prompt cannot reliably stop an agent from posting: it is text, and injected text can argue with it, or the agent can forget it. Removing the ability works. In Evenpost there is no approve tool and no post tool, so no instruction, planted or yours, can make the agent publish.
What should I do with a draft I did not ask for?
A draft you did not ask for should be rejected, not edited into shape. Tap Reject in the Evenpost app and it is archived. Then look at what your agent read just before it drafted, such as a page or a file, because that is the likely source of the instruction.
In Evenpost
Your agent drafts; only you can publish. Connect your agent, or read what Evenpost is.
Updated 6 Nov 2026
Get an invite
An injected instruction ends as a draft you reject, never as a post. Free during the invite-only beta.