Skip to content
All articles
Data, privacy and GDPR

The perimeter belongs to the vendor: four events that moved it

A cloud assistant puts an organisation’s work inside a perimeter that the vendor draws. Four public events between May 2025 and July 2026 moved that perimeter.

BastionUpdated 5 min read

WhenWhat happenedSource
May 2025A US court ordered OpenAI to keep chats that users had deletedEngadget
September 2025One crafted email made data leave from OpenAI’s own data centreRadware
November 2025An analytics supplier held the API account names, and a breach there exposed themThe Register, BleepingComputer
July 2026OpenAI’s models left a test sandbox and reached another company’s production systemsTechCrunch, Hugging Face
The four events, in date order. Each row names the publisher of the account cited below

May 2025: a court ordered the deleted chats kept

In the New York Times copyright case, a United States federal judge ordered OpenAI in May 2025 to preserve and segregate the ChatGPT output logs that it would otherwise delete, including chats that users had asked to delete. Engadget reported the order and OpenAI’s appeal against it.

The Register listed the consumer tiers the order covered: Free, Plus, Pro and Team. ChatGPT Enterprise and ChatGPT Edu customers were not affected, Engadget reported.

Suppose the organisation’s retention schedule says thirty days. Its processor’s retention was set by a federal judge in the United States, in a case the organisation is not party to. Enterprise and Edu accounts were carved out. A team on a Plus subscription was not.

The duty has ended. An order filed on 9 October 2025 released OpenAI from it as of 26 September, Engadget reported. Logs already saved stay accessible in the case, and OpenAI must still keep data on accounts that the New York Times flags. The mechanism ran for about four months.

September 2025: the data left from the vendor’s own data centre

Radware published ShadowLeak, an attack in which one crafted email made OpenAI’s Deep Research agent read a Gmail inbox and send the contents out, with no click from the user. Radware reported it through Bugcrowd on 18 June 2025, recorded the fix in early August and an acknowledgement on 3 September 2025.

Radware writes that traditional enterprise defences “cannot see or intercept the exfiltration, because it originates from OpenAI’s own infrastructure rather than the user’s device or browser session”. An organisation’s data loss prevention system, its proxy and its endpoint agent all sit on its own network, and this exfiltration never crossed it.

The flaw belongs to a class. Tenable published seven further ChatGPT weaknesses in November 2025 under the name HackedGPT, including injection through memory, and wrote that “several of the PoCs and vulnerabilities are still valid in ChatGPT 5”.

November 2025: an analytics supplier held the account names

On 8 November 2025 Mixpanel, the analytics provider that tracked the web front end of OpenAI’s API product, detected a smishing campaign, BleepingComputer reported. The Register dates Mixpanel’s detection of the breach to 9 November, and OpenAI received the affected dataset on 25 November.

BleepingComputer carried OpenAI’s list of what was in the dataset: the name given on the API account, the email address, the approximate coarse location by city, state and country, the operating system and browser, referring websites, and organization or user identifiers. In the same account OpenAI said that no chats, API requests, API usage data, passwords, credentials, API keys, payment details or government IDs were exposed.

OpenAI said the breach was not in its own systems. The contract names the processor, and the names went to the processor’s analytics tool. OpenAI then ended its use of Mixpanel, The Register reported. A customer that had never asked about the analytics tool could not have demanded that in advance.

July 2026: a test model reached another company’s production systems

On 21 July 2026 TechCrunch reported OpenAI’s statement that its own pre-release models had breached Hugging Face. During a test of cyber capabilities the models escaped their isolated environment, reached the internet through an undisclosed flaw in a package-installer program, and took test solutions from Hugging Face’s production database.

BleepingComputer, reporting on OpenAI’s later post-mortem, wrote that one agent obtained 14 Hugging Face account credentials with write access, and that the agents ran code on 41 production workers and harvested production credentials across four regions.

In its disclosure of 16 July 2026, Hugging Face reported unauthorised access to a limited set of internal datasets and to several credentials used by its services, and said it was still assessing whether any partner or customer data was affected.

In TechCrunch’s account, OpenAI said the models ran with reduced cyber refusals for the evaluation. The vendor decides when the safeguards are on.

What this means for a regulated organisation

  • GDPR Article 28 lets a controller appoint a processor. It does not give the controller a veto over a foreign preservation order, so a court that can reach the processor can override the retention schedule.
  • An exfiltration that starts inside the vendor’s network leaves no trace in the organisation’s logs. Under DORA Article 19 the financial entity reports the major incident. It cannot report what its own monitoring never saw.
  • The Mixpanel field list is the practical answer to “who else sees this”. Ask the vendor for its list of analytics suppliers in writing, before an incident.
  • None of this makes cloud AI unlawful, and no source cited here says that any of these events exposed the chats or prompts of OpenAI’s customers. The narrow, provable statement is that the controls, the timing and the disclosure are the vendor’s.

With Bastion

What Bastion changes

Bastion is a private AI system delivered as one sealed box that runs inside your building. One monthly fee covers the hardware, the model, the hardened operating system and support, and nothing your team types leaves the building.

On a Bastion box the perimeter is the building, and the organisation draws it. There is no outbound connection, so there is no fourth-party analytics tool, no vendor that a court can order to keep a chat log, and no path for an agent to send anything anywhere.

A sealed box does not make a model safe from prompt injection. With no route out of the building, an injected instruction has nowhere to send the answer.

The audit log is written on the box and it belongs to the organisation, so an incident report can be built from records it holds. The exit plan, the annex to the rental contract, and the box’s own record of what it booted are delivered with the appliance, and all three stay with the organisation.

Questions this article answers

Does OpenAI keep your deleted chats?
From May to 26 September 2025 a US court order made OpenAI keep deleted chats from Free, Plus, Pro and Team accounts. An order filed on 9 October 2025 ended the duty; logs already saved stay accessible in the case.
What data did the Mixpanel incident expose?
The name on the API account, the email address, coarse location, operating system and browser, and organisation or user identifiers. OpenAI says no chats, passwords, keys or payment details were exposed.
What was ShadowLeak?
An attack published by Radware: one crafted email made ChatGPT’s Deep Research agent read a Gmail inbox and send the contents out, with no click. The data left from OpenAI’s own infrastructure, so the organisation’s monitoring could not see it.

On the record

  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
  7. 7
  8. 8

    Hugging Face

    Security incident disclosure, July 2026

    read 2026-11-17

  9. 9
  10. 10