PRIVACY & CLOUD RISKS
Know where your
business data goes.
Cloud AI adds a provider, its infrastructure, and sometimes connected tools to your data path. Local inference can reduce those external dependencies for the work you keep on your own network.
Sources reviewed 8 October 2026 · Published by Kulper AI Systems
A local client is not always local AI.
Running a coding tool such as Claude Code or Codex on a workstation does not, by itself, keep model inference on that workstation. Check the selected model provider, hosting mode, tools, and network access. Cloud-backed requests are processed remotely; connected services can create additional data flows.
No-training commitments are also different from no-retention commitments. OpenAI says API data is excluded from training by default, while some API data can be retained for abuse monitoring or application features. Anthropic likewise excludes commercial-product inputs and outputs from training by default, with exceptions such as opted-in feedback.
OpenAI API data controlsAnthropic commercial data policyCodex security documentation
Documented incidents. Specific lessons.
These examples describe different failures. A provider incident, a supplier breach, and an agent vulnerability are not interchangeable—and do not establish that every cloud AI product has leaked customer prompts.
ChatGPT exposed some cross-user information
OpenAI reported that a bug exposed some other users' chat titles and possibly the first message of a new conversation. Payment-related information may also have been visible for 1.2% of Plus subscribers active during a nine-hour window. Full card numbers were not exposed; OpenAI fixed the bug.
Why it matters: shared services need effective isolation between customers.
Read OpenAI's incident reportMixpanel analytics data affected some OpenAI users
OpenAI reported that an attacker accessed its analytics supplier, Mixpanel. Limited identifying and analytics information was involved. OpenAI explicitly stated that its own systems were not breached and that chats, API requests, API keys, passwords, and payment details were not exposed.
Why it matters: even metadata creates dependencies on third-party systems.
Read OpenAI's supplier-incident noticeAnthropic documented Claude Code trust-boundary failures
Anthropic described responsibly reported Claude Code vulnerabilities and a controlled internal exercise involving unauthorized data access. Its report discusses fixes and stronger containment. These are disclosed vulnerabilities and test findings, not evidence of a mass customer-data breach.
Why it matters: an agent's permissions and connected tools matter even when it runs locally.
Read Anthropic's containment reportReduce what needs to leave.
For a local workflow, keep the model runtime, document index, and stored conversations on your infrastructure. Set explicit access rules for external tools, remote administration, logs, and cloud fallbacks. Local hosting still needs secure accounts, patches, backups, and sensible agent permissions.
See our planned management approachSTRAIGHT ANSWERS
Common questions
Do cloud AI providers train on every business prompt?
No. OpenAI says API data is not used for training by default, and Anthropic states the same for its commercial products. Exceptions, retention, feedback settings, and product-specific terms still need to be checked.
Is no training the same as no retention?
No. A service can exclude your content from model training while retaining it for operations, safety, or product features. Check the exact product, feature, region, and contractual controls.
Does running a model locally prevent every breach?
No. Local inference reduces the need to send inference data to an external provider. It does not remove risks from compromised accounts, malicious documents, connected tools, poor permissions, or unpatched systems.