A security start-up called Cathedral penetrated internal systems of the AI company Anthropic as part of a bug bounty programme.The route in did not lead through the company's core systems, but through third-party software.AI agents were used for this, i.e. programmes that independently plan tasks, call tools and derive their next steps from the intermediate results.
Anthropic confirms that sales, HR and IT data were accessed, but disputes the description of the full extent.Regardless of how far the access actually reached in detail, the core of the story remains: even a company that itself works at the forefront of AI systems is vulnerable through its external dependencies.
Why this is relevant for completely ordinary web projects
The constellation is not exotic.Every mid-sized company today operates a chain of systems that talk to each other: a CMS, a CRM, a ticket system, chat tools, marketing automation, accounting, single sign-on.These systems are connected via interfaces and access tokens, i.e. technical keys with which one service identifies itself to another.
These tokens are often granted generously.They are created once for an integration, given far-reaching rights as a precaution and then remain valid for years.Anyone who compromises such a connection is not standing at the front door, but already in the hallway.
The case described also makes clear that the vulnerability does not have to lie in your own code.A third-party provider with access to your systems effectively becomes part of your attack surface.Its security level is your security level.
What AI agents change about the situation
Attacks on software supply chains have existed for years.What is new is the speed and persistence with which they can be carried out.An agent can go through hundreds of integrations, subdomains and configuration files, combine clues and derive the next step from them, without a human evaluating every branch individually.
This shifts the ratio of effort to yield.Attack paths that were previously too laborious to pursue against a mid-sized company become economically viable.At the same time, the same possibilities arise on the defence side: checking your own landscape can also be automated.
There is also a second point that is still underestimated in many projects.Anyone who uses agents themselves, for example for content maintenance, for analyses or in support, gives an automated system access to data and tools.Every such permission is a possible lever if the system is manipulated, for example through prepared content that the agent processes.
Concrete starting points in CMS and web projects
The lessons from the incident can be translated into manageable measures that can also be implemented without your own security team:
- Inventory of integrations: Which external services are connected to the CMS, the shop, the intranet?Who set up these connections, and are they still in use?Unused connections are the cheapest security gain there is.
- Consistently limit rights: A token for an import needs write rights on a few data records, not on the entire system.Reading and writing should be separated, administrative rights the exception.
- Tokens with an expiry date: Access keys that are renewed regularly lose value for attackers.This requires that it is documented where which key is used.
- Logging with evaluation: Access via interfaces should be traceable.What matters less is that logging takes place, but that conspicuous patterns are noticed by someone.
- Keep an eye on dependencies: Extensions, libraries and packages in the project should be kept up to date.In TYPO3 projects this affects both the core and the extensions used.
- Treat agent access like user accounts: An agent gets its own, clearly delimited account with its own rights and its own logging, not a shared administrator account.
Think organisationally rather than purely technically
Many of these points are not programming tasks, but questions of responsibility.Who decides whether a new tool is connected to the CMS?Who checks which data flows out in the process?Who notices that a service provider has left the project and their access still exists?
In practice, gaps rarely arise from a wrong line of code, but from no one being responsible.A concise, up-to-date overview of systems, connections and responsible parties is therefore often more effective than another technical layer of protection.
It is also helpful to plan security as part of operations and not as a one-off project step.An annual review of integrations and permissions, with a fixed date and a clear result, costs little time and prevents an confusing landscape from building up over the years.
Assessment
For web and software projects, the perspective shifts: your own application is only as secure as the weakest connection into its environment.Anyone who documents interfaces, access rights and external dependencies and reviews them regularly significantly reduces the risk without major investments.And anyone integrating AI agents into their own processes should scope their permissions as tightly as the task allows from the outset.