Cognition published a security guide arguing that established defensive practices remain important as AI accelerates technical work. The recommendations span corporate identity, developer devices, source code, third-party dependencies, and infrastructure monitoring.
The guide emphasizes reducing production credentials on developer machines, auditing extensions and agent integrations, pinning dependencies, and preserving useful audit logs. It also recommends giving code-review tools an explicit threat model. The unifying idea is that an agent inherits the capabilities of the environment in which it runs, making that environment a central part of the organization's security design.
Start with a map of access
A development workstation often connects source code, package registries, cloud services, and deployment systems. An agent that can run a shell may reach several of those through existing configuration. The organization should know which connections are necessary for a task and which are simply present because of historical convenience.
A useful first exercise is to trace one ordinary change from checkout to deployment. Record the identities used at each stage, the secrets available, and the point where production authority becomes available. This can reveal a broad credential that is unnecessary for most development work.
Moving execution to a remote environment can make isolation easier to manage, but the label alone does not establish a safe boundary. The remote worker still needs restricted access, an update process, and a way to remove state when the task ends. The same questions apply wherever the code runs.
Make maintenance part of the workflow
A dependency inventory becomes useful when someone owns the response to a change. Pinning a version improves reproducibility, but a pinned vulnerable package still needs to be updated. A deliberate review path should connect notices, testing, and rollout rather than leave the inventory as a static document.
Logs need similar ownership. Decide which events would help reconstruct an incident and verify that those records are available independently of the affected application. A log that exists only inside a disposable worker may vanish precisely when it is needed.
For teams adopting agents, the practical opportunity is to improve these foundations during the rollout. Define a small execution role, give it the access required for one workflow, and expand only when a demonstrated need appears.
Cognition's guide is most useful as a prompt to examine the real development chain. Stronger controls around identities, dependencies, and evidence can benefit both human developers and agents, because both rely on the same underlying systems to turn code into production behavior.
Have a correction or a useful perspective?
Get in touch with the publication.


