The real workflow
Where OpenAI Codex enters the work
An agency coding agent can cross client boundaries when repositories, terminals, credentials, caches, or sessions are reused between engagements.
OpenAI Codex can work locally or in cloud environments with repository files, commands, patches, network controls, approvals, plugins, and connected developer workflows.
Long-running or cloud tasks can continue across multiple steps inside the permissions and integrations granted at launch.
Codex behavior depends on the environment, sandbox profile, approval policy, network access, connected services, and task scope. A protected default can still be widened by explicit authorization.
The presence of this path does not prove an incident. It identifies the boundary that should be checked before more sensitive context or authority is added.
Tool-specific boundary
Inspect the real access points.
What may carry context
local repositories and worktrees
commands, patches, tests, and tools
cloud repositories, plugins, MCP servers, and network access
Settings to verify
Sandbox and approval profile
Writable roots and network policy
Repository, plugin, MCP, and cloud connections
Why this context matters
The consequence for agencies
Agency risk compounds when staff, contractors, shared tools, and reused credentials create paths between otherwise separate client environments. In this case, at work, weak approval boundaries can affect customers, communications, infrastructure, financial operations, permissions, and auditability across multiple connected systems.
Autonomy changes the failure mode. A bad answer can be ignored; a bad action may already have changed a file, sent a message, altered access, spent money, or affected production before someone notices.
Every client remains isolated, access is attributable to a named operator, and the agency can deliver consistent evidence without revealing another client.
Context decision
Three questions before adding access
Can this operator or tool reach any repository, mailbox, drive, cache, token, or transcript belonging to another client?
Are credentials and AI sessions issued per client and person rather than shared across the agency?
Can the agency deliver useful proof to this client without including another client's names, paths, findings, or configuration?
Evidence goal: Create a separate client evidence record covering operator identity, workspace isolation, credentials, approved systems, review history, and delivery status.
A repeatable review
Four steps, no sensitive data required
- 1
Write down the exact OpenAI Codex account, workspace, project, device, and connected service used in this workflow.
- 2
Set a bounded objective, inspect the environment and approval profile, and define merge or external-write stop points.
- 3
Assign the decision and next review to the client service owner or agency security lead; do not leave the access boundary as an unwritten assumption.
- 4
Require review before repository publication, issue creation, deployment, or any action outside the worktree. Record the result without copying private content or raw credentials into the report.
Controls to apply
Reduce access before adding trust
Require review before repository publication, issue creation, deployment, or any action outside the worktree.
Keep consequential actions on ‘always ask’ or equivalent unless a narrowly scoped policy justifies otherwise.
Set limits for money, recipients, repositories, branches, destinations, records, and time windows.
Provide rollback, revocation, and a tested stop mechanism before background execution.
Decision rule
Know when a formal baseline is justified
Text-only assistance does not create autonomous-action risk. When the tool can change the outside world, formalize approval and evidence before increasing speed or scope.
CapitalGuard is relevant when the workflow includes repositories, recurring private work, credentials, connected systems, commands, or evidence that must be shared with another person. It does not inspect this account from the page or guarantee that an incident cannot occur.
Primary references