Skip to content
S&L Labs

Writeups

Exit readiness is a security property

A locked account, a failed region or access by a foreign jurisdiction hits a system as hard as a bug in the code. How I inventory and assess a system's dependencies on its vendors, and what is different with agents.

  • Availability
  • Lock-in
  • Architecture
  • Simon Liedtke

My analysis, written up with AI.

A system can go down without anyone exploiting a vulnerability. The account at the cloud provider gets locked, a region fails, an identity service is unreachable, or a foreign jurisdiction demands access to data. For users, the result looks the same as after an attack: the service is gone, or the data is no longer under your own control.

Availability is one of the three protection goals of information security, alongside confidentiality and integrity. Even so, reviews rarely treat vendor dependency as a security topic; it usually ends up filed as a purchasing or strategy question. I treat it like any other attack surface: I inventory the dependencies, ask for each one what happens if it is locked or fails, and record which ones stay bound on purpose. In this sense, exit readiness is a property of the architecture that can be checked.

1. Four ways a vendor becomes a risk

The account gets locked. In May 2025, AP reported that the chief prosecutor of the International Criminal Court had lost access to his Microsoft email account following US sanctions (Quell 2025). Microsoft president Brad Smith denies having ended or suspended services for the court. The company does, however, acknowledge that the sanctioned prosecutor was cut off from Microsoft services (Krempl 2025a, 2025b). Regardless of how the case played out in detail, it shows the kind of risk: a decision made outside your own system takes away one of its central functions. A lockout can also come about in more mundane ways, through a failed payment, an automated abuse flag or a dispute over terms of service.

A service or region fails. On October 20, 2025, a fault in the automated DNS management of DynamoDB largely took down the AWS region us-east-1 for hours (Amazon Web Services 2025). On October 29, 2025, a faulty configuration change to Azure Front Door disrupted Microsoft 365 and Entra ID, among others, for about eight hours (Microsoft 2025). A system that runs in only one region, or whose sign-in depends on a single identity service, goes down with it in such cases, even if its own code is free of errors.

A foreign jurisdiction demands access. For data held by vendors based outside the EU, the question is which authorities can demand access and on what basis. The legal assessment belongs with lawyers and data protection officers. For the architecture, a simpler question counts: which data and keys sit with which vendor, and could they be moved?

Keys and access are concentrated. In my writeup on the LLM gateway I described why an LLM gateway that bundles all provider keys is an attractive target. The same applies to a cloud account through which identity, key management, backups and production all run. Whoever takes over or locks that account hits everything at once.

2. Sovereignty is a spectrum

The desire for independence has grown measurably. In Bitkom’s Cloud Report 2026, 85% of the companies surveyed consider Germany too dependent on US cloud providers, up from 78% the year before. 59% name lock-in, meaning difficult data export and difficult migration, as the biggest obstacle to switching providers. Only 12% would accept 10 to 20% higher costs for a cloud that processes data exclusively in Germany, and 43% see no equivalent European alternative for their needs (Bitkom 2026).

The market is moving slowly accordingly. According to Synergy Research, the share of European providers in the European market for cloud infrastructure services has held steady at around 15% since 2022 (Synergy Research Group 2025). At the same time, the large providers offer their own variants. Since January 15, 2026, AWS has operated a European Sovereign Cloud with its own parent company and three German GmbHs (Amazon Web Services 2026). Whether such a variant is enough for a given purpose is a trade-off, and the answer is rarely a plain yes or no.

For teams that build their own software, this suggests a pragmatic stance. Asking which vendor is sovereign enough leads nowhere concrete. The useful question is what happens if a particular vendor fails or can no longer be used, and how long it takes until the system runs again. For each dependency there is a concrete answer to that, and the answer can be improved.

3. Inventorying dependencies

I start with a list of all services without which the system does not run or data gets lost. These categories are typical:

  • Identity: identity provider, single sign-on, MFA
  • Compute and managed services: container platform, functions, workflow services
  • Databases, queues and object storage
  • Keys: KMS, secrets management, certificates
  • CI/CD, code hosting and container registry
  • DNS, CDN and email delivery
  • Observability: logs, metrics, alerting
  • for AI features: LLM API, embedding model, vector database, third-party MCP servers

For each entry I answer four questions:

  1. What happens if it is locked or fails? Does the whole system go down, only one feature, or nothing at all because there is a second route?
  2. How does the data get out? In which format, through which interface, in how much time? An export that nobody has ever tried is an assumption.
  3. Where are backups and keys? A backup in the same account as production helps against deleted data. It does not help against a locked account. The same applies to keys that exist only in the vendor’s KMS.
  4. Is there a documented alternative? By this I mean a description of where the system would move in an emergency and what that would take. A running parallel setup is not required.

An excerpt from such an inventory for a SaaS application with an agent feature could look like this:

Dependency If locked or failed Getting data out Decision
Identity provider (SaaS) no sign-in, system down users exportable, passwords not emergency access for admins outside the IdP
Managed Postgres system down standard dump, tested bound on purpose, backup in a second account
Cloud provider’s KMS encrypted data unreadable keys not exportable data keys also backed up offline
LLM API agent feature fails nothing to migrate second provider configured, evals for both
Vector database (SaaS) search fails vectors exportable, source texts held by us rebuild from source texts documented

The “Decision” column is the purpose of the inventory. “Bound on purpose” is a legitimate answer there. An abstraction over every service costs time, prevents the use of the features that make a vendor attractive in the first place, and brings its own sources of error. I recommend it where a failure or lockout hits the entire system and switching would take weeks today. For the rest, a tested export and a note on what to do in an emergency are often enough.

4. What is different with agents

With agents, swapping the model is considered solved, because many frameworks support several providers. In practice, more depends on the model than the API call. Prompts, tool descriptions and the output format are tuned to one model, and a different model behaves differently with the same instructions. An agent is only exit-ready once there are evals that show whether it still does its job with the second model.

A vector database has a less obvious tie. The stored vectors belong to the embedding model that produced them. If you change the embedding model, you have to recompute the entire collection. That only works if the source texts and metadata are fully available outside the vector database.

Third-party MCP servers are dependencies with write access. If such a service fails or changes its behavior, that affects the integrity of the actions the agent performs through it, in addition to availability. In the writeup “MCP attack patterns, mapped to OWASP Agentic 2026” I described how tool descriptions can be pinned. For the exit question, add whether there is a second route for every critical tool, or whether the agent then declines the task cleanly.

Finally, the account at the model provider. A locked or compromised account stops all agents at once. Separate projects per application, limits and a second provider with its own account contain this.

5. What is out of scope

Since September 12, 2025, the Data Act has required cloud providers to enable their customers to switch. Until January 12, 2027, they may only charge reduced switching fees that do not exceed their costs directly linked to the switch. After that, they may not charge any switching fees at all, including fees for data export (European Union 2023, Art. 29). That is a good occasion to check your own exit readiness now. Whether a contract or a vendor meets the legal requirements is a question I do not answer. The same goes for the CLOUD Act, the EU’s proposal for a Cloud and AI Development Act and the GDPR. I also do not recommend vendors. My contribution is the technical map: which dependencies exist, what a failure means and what a switch would cost.

What I check here

Together with the team, I build the inventory of dependencies and answer the four questions from section 3 for each one. I check whether backups and keys are stored outside the account they are meant to protect, whether there is emergency access that does not depend on the identity provider, and whether exports have ever been tested. For agents, this includes the evals for a second model, the source texts behind the vector database and the dependency on third-party MCP servers. The result is a list of dependencies, each with a decision: bound on purpose, safeguarded or to be rebuilt.

Sources