Anthropic's 30-day retention rule prompted a new answer: OpenAI and Anthropic now keep AI safety logs in the customer's own cloud. Control goes up, and so does responsibility. What it means for organisations in Indonesia.
Tim MarkasDevTechnology Partner, Jakarta
When your team pastes a contract, customer data, or source code into an AI model, there is a question that rarely came up in procurement meetings: once the answer comes back, is that prompt still stored, where, for how long, and who holds the keys? Over mid-2026, that question moved from a footnote in the terms of service to a point of negotiation between the largest AI model providers and their enterprise customers.
For leaders and operations managers in Indonesia, the shift matters because it moves part of the responsibility to the customer. More control comes with more work.
Anthropic announced that from 9 June 2026, prompts and outputs of what it calls covered models, meaning Mythos-class models and similar models it designates later, are retained for 30 days for safety work, on every platform where those models are offered. According to Anthropic's Privacy Center, the change affects organisations that previously used zero data retention (ZDR), an arrangement in which prompts and responses are not stored after processing. The reason given: some attacks only become visible when many requests are analysed together, such as jailbreak attempts that send hundreds of slight variations of a prompt.
The market pushed back hard. Fortune reported broad resistance from companies worried about their data and intellectual property leaking. Two responses followed, with a similar pattern:
| OpenAI: ZDR with PSP | Anthropic: EFS | |
|---|---|---|
| Where data lives | Customer cloud storage, per region | Customer cloud account (S3, Azure Blob, GCS) |
| Keys | Double encryption; outer layer uses customer keys (EKM, recommended) | Customer-managed encryption keys (opt-in) |
| Retention | At least 30 days, set by a lifecycle rule | Rolling monitoring window |
| Findings | Bounded signal to OpenAI; customer investigates | Sent to the customer's team |
| Status | Phased since 22 September 2026 | Phased from autumn 2026 |
Both answer the same dilemma: the provider needs interaction history to detect misuse, while customers, especially in regulated sectors, do not want that history sitting on someone else's servers.
This is where operational detail starts to matter. OpenAI's technical guide requires customers to register and validate storage for every project and data residency location, keep encrypted records for at least 30 days, keep storage permissions and key access correctly configured, fix configuration problems when notified, and respond to notices about safety concerns. The guide also warns that a "Validated" status records one successful check, not ongoing storage health. One more detail that is easy to miss: PSP is enabled per project, so all API traffic in that project follows the same policy.
Anthropic does not charge for EFS, but the customer's cloud provider still bills for storage, reads, writes, and egress. A CISO quoted by Fortune put it well: you are trading control for responsibility. That is why some companies take a hybrid route: frontier provider models for some work, open models in their own environment for the most sensitive data.
Many organisations in Indonesia use AI through individual accounts or team subscriptions without ever checking the retention settings. Three points are worth noting.
First, the responsibility stays with you. Under Law No. 27 of 2022 on Personal Data Protection (the PDP Law), Article 51 says a processor must process data on the controller's instructions, and that processing remains the controller's responsibility. If an AI provider processes your customers' personal data, your organisation is still the one that has to answer for it.
Second, "stored in your cloud" does not necessarily mean "stored in Indonesia". For Azure, OpenAI's guide asks for an approved US or EU storage region; for AWS without data residency, the recommended region is us-west-1. That means the data still leaves the country, and Article 56 of the PDP Law applies: the receiving country must offer equal or higher protection, or there must be adequate and binding safeguards, or the data subject's consent.
Third, misuse signals need a recipient. If findings are sent to your team, it must be clear who reads them, how fast, and what happens next. This follows the principle we covered in AI agents need access limits.
Practical steps you can start this week:
At MarkasDev we split this work into three phases. Diagnose maps data flows, classification, and compliance gaps before budget is committed. Deliver builds the gateway, project separation, and storage under your keys, in milestones with written acceptance criteria, as part of our AI implementation service. Operate keeps storage validation, keys, and signal handling in check under an SLA.
If your organisation already uses AI but is not sure where its data goes, request a Diagnose through our contact page. We start from the tools your team is using today.
Cisco and Grafana 2026 surveys: observability spend is rising, verbose logs are the top cost driver, and telemetry rarely informs business decisions. Seven steps to fix it.
A WhatsApp message and an AI-cloned voice led Fideuram to wire 95 million euros abroad. The defence is not a detector but the payment approval flow. Seven steps for finance teams.
