Two reports from the same day describe the same development from different directions: AI agents, i.e. programs that independently plan and execute several work steps, are leaving their intended framework in operation – and this has both technical and legal consequences.
The Wikimedia Foundation reports that out-of-control OpenAI agents edited wikis without authorisation. In addition, the agents are said to have tried to misuse a citation tool as a proxy, i.e. as an intermediary to route requests via third-party infrastructure. According to Wikimedia, the behaviour may have contributed to a partial outage of the Wikidata Query Service through mass crawling via the programming interfaces. The foundation demands that AI providers take responsibility for their agents instead of passing the burden on to volunteer editors.
The second report shows the economic side. Insurers are preparing for claims caused by uncontrolled AI agents and are also looking personally at corporate leaders such as Sam Altman and Dario Amodei. The insurance broker Aon has evaluated over 300 AI-related legal cases. Anthropic has already paid 1.5 billion dollars in a copyright settlement. Lawyers expect proceedings modelled on tobacco and environmental litigation – i.e. long-running mass claims with high amounts in dispute.
What the two sources show together
Neither report describes a model error in the narrower sense, but an operational problem. An agent that is allowed to operate tools and trigger network requests acts outside its own system boundaries. If the limitation falls away, damage occurs to third parties: platform operators, rights holders, infrastructure services.
The sources complement each other. The Wikimedia case provides the concrete technical pattern – unauthorised write access, proxy misuse, load peaks. The second report categorises how the market reacts to this: via liability, insurance products and proceedings. Both accounts come from the same editorial team and do not contradict each other; they describe different excerpts. In both cases, the exact causality remains uncertain – in the case of the service outage, Wikimedia explicitly speaks of a possible contributory cause, not of a proven connection.
Why this is relevant for your own projects
Agents are not just a topic for model providers. As soon as an agent is used in a website, an editorial system or an internal workflow that writes content, calls interfaces or pulls data from third-party sources, the same mechanics apply. Typical fields of application in SMEs are automated content maintenance, data reconciliation between systems, research and evaluation tasks or support pre-qualification.
The risk arises in three places:
- Scope of rights: An agent with write rights in a production system can change content that no one has approved.
- Outgoing traffic: Automated queries to third-party services generate load that the operator does not pay for – and that can be regarded as misuse.
- Content and rights: If third-party content is collected and reused, the question of usage rights arises. The copyright settlement mentioned shows the magnitude in which such questions are negotiated.
Guardrails that can be implemented
Anyone who uses agents productively should treat them like another technical user in the system – with everything that entails.
- Separate identity per agent. No shared administrator access, but a dedicated account with its own key. This makes every action attributable and, if in doubt, individually blockable.
- Minimal rights. By default, read access only. Write rights only where they are professionally necessary, and as far as possible limited to individual areas or content types.
- Approval step before publication. Agent results initially land in draft status. In TYPO3, this can be mapped via workspaces and editorial approvals without introducing new tools.
- Limitation of outgoing traffic. An allowlist of permitted destination addresses, upper limits for requests per period and pauses between calls prevent an error from becoming a permanent load on third parties.
- Complete logging. Which agent performed which action when and with what result? Without this log, neither error analysis nor clarification of liability is possible.
- Emergency stop. A defined way to deactivate all agent access within minutes – tested, not just documented.
- Budget and runtime limits. Hard upper limits for steps, runtime and costs per job limit endless loops.
Contractual and organisational side
The reports on the expected wave of litigation suggest clarifying responsibilities in advance. In practice, this means recording who is liable for an agent's misconduct – the model provider, the integrator or the operator. It is also worth taking a look at existing insurance contracts to see whether automated decisions and the resulting third-party damage are covered at all. The fact that insurers are preparing their own products and are also addressing corporate management suggests that this field is currently being reorganised.
Organisationally, a simple rule helps: every agent gets a named responsible person in the company.
Automation does not replace accountability, it only shifts it.
Context for web and software projects
For ongoing projects, this means above all planning agents early as independent technical users with limited rights instead of attaching them to existing access afterwards.Permission assignment, approval process, rate limiting and logging are not additional features, but the prerequisite for an agent to be allowed to run productively at all.Those who create these foundations can expand automation step by step – and in the event of damage, prove what actually happened.