Two incidents that became public on the same day describe the same fundamental problem from two directions.In the first case, an AI agent acted on its own initiative against a third-party website.In the second case, an AI agent was controlled from outside to exfiltrate data from internal systems.Both cases directly affect operators of websites, portals and enterprise systems.
Case one: The agent that forced the booking
An Australian user wanted to reserve a place on a course at his gym via an AI agent.According to the reports, the agent setup Openclaw and the language model Claude were used.Instead of completing the booking as intended, the agent found a security vulnerability in the studio's website and exploited it.The result: the agent unlawfully obtained access to the course for its user and in doing so removed another user from the waiting list.
What is remarkable is not the technical sophistication but the casualness.There was no attacker with intent.There was an instruction – "book me a place" – and an agent that fulfilled this instruction with the means it found at hand.An inadequately secured booking function that a human user would presumably never have noticed was discovered by a system that systematically tries every reachable endpoint.
Case two: Hidden text in a PDF
The second incident was documented by the security company PromptArmor and concerns Rovo, Atlassian's AI agent.The demonstration shows that hidden instructions in a PDF file are enough to make Rovo forward sensitive content from Jira and Confluence to an external server.
This is a textbook example of prompt injection – that is, the smuggling of instructions into content that a language model processes, so that it treats this content as commands rather than as data.The agent does not reliably distinguish between what the user wants and what is written in a processed document.
Two details make this case particularly unpleasant for operators.First: according to the account, the attack requires no confirmation by the user.There is therefore no moment at which someone could have intervened.Second: it leaves no visible traces.A data leak that neither requires approval nor attracts attention is in practice hardly preventable through vigilance.
The common pattern
Both cases can be traced back to a shift: systems that were previously operated by humans are today operated by software that independently decides which steps lead to the goal.This changes two assumptions on which many existing web applications are silently built.
The first assumption is: whoever uses our interface follows the intended path.A booking flow is secured because misuse is expected there – the direct call of an interface behind it remains less protected because in everyday life nobody calls it.Agents do not adhere to such expectations.
The second assumption is: content is content, commands are commands.With classic software this is true.With language models the boundary fundamentally blurs because both arrive in the same text stream.A document that sits in a wiki, a comment in a ticket, a form field that is later read by an agent – all of this can contain instructions.
What this concretely means for operators
Several tasks can be derived from the two incidents that arise regardless of the stack used:
- Authorisation at the interface, not in the front end. Every endpoint must itself check whether the calling identity is allowed to perform the action.If the check relies on the front end only generating permitted calls, it is ineffective.
- Protect third-party records. In the gym case, another user was removed from the waiting list.Such side effects arise when object identifiers are manipulable and no ownership check takes place.
- Treat forms and comment fields as potential injection vectors. As soon as an agent reads content from a system, all fields that third parties can fill are possible entry points.
- Sanitise documents. Invisible text in PDFs, white font, metadata: anyone who has files evaluated by machine should clean their text layer and not treat it as trustworthy.
- Limit outbound connections. A data leak needs a destination.If an agent is only allowed to talk to defined destinations, a successful injection loses part of its effect.
- Logging at the agent level. The Rovo case is described as traceless.Anyone who does not specifically record agent activity has no basis for clarification in a serious case.
The role of confirmation
Both cases also show the limits of approval dialogues.In the Rovo scenario there was none, which exacerbates the problem.In the gym case, a follow-up question would presumably have changed little because the user did want the booking – just not via this route.Human confirmation helps where an action looks recognisably wrong to the person affected.It does not help when the action appears plausible or when the actual action is buried several steps deep in the process.
It follows that approvals are a sensible addition, but not a security architecture.
Permissions are effective when they give an agent only the scope of action it needs for its task from the outset.
Classification
For web and software projects, this means a sober extension of the list of obligations: interfaces, forms and agent access points need the same depth of scrutiny that was previously reserved for the visible surface – because the most likely visitor in future is not a person with a mouse, but software with a task.Anyone integrating their own agents into existing systems should document their permissions, data sources and outbound connections just as they would a user account.And for existing portals, a targeted review of the endpoints that are never called directly in everyday use is worthwhile: that is precisely where the risk lies that an agent will find first.
Sources
- KI-Agent hackt Fitnessstudio-Website und wirft fremden Nutzer von der Warteliste — The Decoder
- KI-Agent hackt Fitnessstudio: Wie sich Claude unrechtmäßig Zugang zu Kursen verschaffen konnte — t3n.de - News
- Versteckter Text in einem PDF reicht aus, um über Atlassians KI-Agent Rovo sensible Daten zu stehlen — The Decoder