
The last two weeks were about AI finding flaws, first in PHP and then in the Linux kernel. This week is about where AI data sits. On 9 November Notion told customers that AI chat transcripts from fewer than 1% of its workspaces were in a dataset accessed by attackers at Fireworks AI, one of its AI suppliers.
A day later the hosting company Railway published its full report on an October intrusion in which, it says, AI provider keys were the main target.
Neither company sells to construction. Both reports matter to PMs, cost teams, contractors, designers and clients because firms tell clients that their AI use is 'zero retention', or that project data never leaves approved tools. These two documents show how much sits underneath that sentence.
What Notion and Fireworks have said
The incident unfolded over two weeks, with Notion initially believing that only synthetic testing data had been accessed. It was only several days later that Fireworks flagged the possibility of customer data exposure. The subsequent investigation revealed that some Notion AI chat transcripts were part of the dataset.
It started at Fireworks: A security incident at a third-party cloud provider on 23 October gave an unauthorised party access to Fireworks' development machines the following day. Its security update says the resulting activity led to additional customer data exposure.
Notion initially saw little cause for concern: Fireworks contacted Notion on 2 November about a small amount of accessed data, which Notion identified as synthetic testing data. Notion disconnected its systems the next morning and deleted its data from Fireworks on 4 November.
The picture changed: On 6 November, Fireworks warned that customer data might also have been accessed. It confirmed the final dataset on 8 November, and Notion began contacting affected customers the following day.
What was exposed: The dataset included Notion AI chat transcripts collected through Insights, a feature that helps customers understand their AI usage. Notion says most transcripts were not labelled with individual users or workspaces.
Notion has since disabled Insights and stopped serving open-weight models through Fireworks. It also advises affected customers to treat the data as exposed and rotate any passwords or API keys they may have pasted into Notion AI. Its advisory, updated on 12 November, has the full details.
Where 'zero data retention' stopped
Fireworks also says: 'We have not detected any impact to Inference, Training SDK and Nexus, which are all zero data retention (ZDR) by default.' On its own terms the promise held. Its documentation states that 'Zero Data Retention does not apply to certain services that require data storage, including Batch API and Responses API.' Lin Qiao, Fireworks' CEO, was more specific in a post on X quoted by VentureBeat, listing 'Inference (excluding batch inference)' among the products with no detected impact. The words in brackets are the part a buyer needs to read.
How Notion's transcripts came to be stored is still unclear.
Neither company has said which Fireworks feature stored Notion's transcripts, and nothing in the disclosures shows the policy would have stopped the intrusion.
What the two advisories do establish is that a feature inside an everyday work tool sent chat content to a supplier, and the supplier held it somewhere its default did not cover.
Retention terms also move with the model.
Anthropic's help centre says that for its Mythos-class 'covered models', 'Prompts submitted to, and outputs generated by, covered models are retained for 30 days to support our safety work, on every platform where these models are offered.'
The policy has applied since 9 June 2026, and only some zero-retention customers are eligible for an exception under what Anthropic calls 'Enterprise Frontier Safeguards'. A statement made to a client last year may not describe the model a team uses today.
Railway: the keys were the target
Railway hosts applications for developers, who store API keys and other secrets with it. Its incident report of 10 November says an attacker chained several exploits on 20 and 21 October and moved across its internal systems. Railway can validate that 'at most 1.85% of workspaces, about 100,000, could have had any of their environment variables decrypted', and it calls that figure an upper bound.
Unauthorised spend on the stolen keys 'was estimated to reach ~$740,000'. Railway says it is working with AI providers on refunds and will cover what they decline to.
Railway's own reading is wider than one company: 'We assess this to be one of a series of attacks on platforms that hold AI keys.'
Benjamin Read and Rami McCarthy of the security firm Wiz counted at least nine AI developer platforms disclosing incidents in October. 'All public reports highlight targeted exploitation, explicitly hunting stored environment variables, with a focus on those granting AI access,' they wrote.

The scale each company reports. Source: Notion security advisory (12 November 2026) and Railway incident report (10 November 2026).
What UK guidance already expects
New rules are not required for any of them. The ICO's guidance on controller and processor contracts says a processor should not engage a sub-processor without the controller's prior written authorisation and that 'the processor is liable to the controller for a sub-processor's compliance with its data protection obligations'. That guidance is under review following the Data (Use and Access) Act.
The NCSC's supply chain guidance, last reviewed in October 2025, observes that 'many companies lose sight of their supply chains'. A meeting-notes tool with an AI summary button has a supply chain of its own.
Takeaway
A retention promise covers the features its author had in mind. Before repeating one to a client, find out which features those are and who else handles the data.
Ask each software supplier for its sub-processor list and which sub-processors sit behind each AI feature, including usage dashboards and other analytics.
Read the retention clause for what it leaves out: batch jobs, stored responses, logs, evaluation and training features, and safety retention on newer models.
Find out where your API keys for AI services are stored, whether any sit in shared hosting environments, and whether each has a spend limit and an alert.
Check the wording in your bids and client AI statements. If it says 'zero retention', it should name the tools and features it applies to.
Supplier advisories rarely reach a project director's inbox, yet they decide what an AI statement to a client is worth. We read them every week. Subscribe to the Project Flux newsletter to get the briefing on Monday.
Links and Stuff
All content reflects our personal views and is not intended as professional advice or to represent any organisation.

