Fireworks AI now says its October activity was two separate security incidents, and that the second one, which ran from 23 October to 30 October 2026, reached account and billing information for all accounts created before 30 October. Uber has filed breach notices with several U.S. state attorneys general over driver's license data accessed through a service provider, and an Uber spokesperson told TechCrunch the vendor was "Fireworks.ai." Uber's entry on the Oregon Department of Justice data breach list records 149,346 people affected; TechCrunch reports this is the total Uber notified. Cursor said on 13 November that an attacker used an access key it had issued to Fireworks, which Cursor says was "exposed by Fireworks." Fireworks' bulletins name none of these companies, and Fireworks has now published indicators of compromise. This post reflects sources checked as of 18 November 2026, 15:02 EST.
Notion says the incident exposed Notion AI chat transcripts from fewer than 1% of Notion workspaces. Per Notion's advisory (updated 12 November 2026, with a further update posted at 19:15 EST that day, by Rescana's conversion; Notion's page dates it 13 November), the data came from the dataset behind Insights, which shows administrators how their teams use Notion AI. Notion says that for the vast majority of impacted workspaces the transcripts were "unattributable": they were not labelled with who wrote them or which workspace they came from, and did not contain tool call results. Notion states that "this was not a breach of our systems." It has disabled Insights and stopped serving open-weight models through Fireworks.
Technical Information
Fireworks: two incidents, ending 30 October
Fireworks' 17 November bulletin entry says the activity in its earlier 10 October and 9 November entries "relates to two separate security incidents":
| Incident (per Fireworks) | Window (2026) | How it happened | What was accessed | Investigator |
|---|---|---|---|---|
| First | 3 August to 7 October | Stolen Fireworks credentials plus "a vulnerability in one of our services" | Platform misuse; for a small number of customers, secrets, model files and datasets | Unit 42; forensics completed |
| Second | 23 October to 30 October | Development machines hosted at a third-party cloud provider that had its own security incident, used "to reach part of our production systems" | Account and billing information for "all accounts created before Oct 30th"; for a small number of customers, datasets, models and/or secrets | AprosSyber; investigation under way |
Fireworks says both incidents are "contained and remediated." Per the 9 November entry, billing information means addresses, invoice and payment history, account balance and credit usage; Fireworks does not store payment card numbers.
The 25 October line is a detection statement, not the end date. The 9 November entry says that on 24 October the attacker used the cloud provider incident to reach Fireworks development machines and run an unauthorized workload, that Fireworks detected this the same day, and that "as of October 25, we have not detected further evidence of successful unauthorized access by the threat actor on the Fireworks platform customers use." That sentence is scoped to one surface (the customer-facing platform) at one point in the investigation. The later 17 November entry gives 30 October and production access. Fireworks has not explained the difference, and we do not speculate about it. For operations, use the later date.
The first entry, dated 10 October, said Fireworks identified "unauthorized use of internal Fireworks credentials" on 6 October and that environment variables in evaluation jobs, "including API keys and tokens stored as secrets," were captured in certain accounts. Its 9 November entry, written before Fireworks split the activity into two incidents, says the evidence "indicates links to a cybercriminal threat actor group, which is financially motivated" and aims to sell tokens from unauthorized LLM access to illicit routing gateways ("LLMJacking"). It has not named a group; TechCrunch reports no known group has taken credit. This post makes no further attribution.
Uber: valid vendor credentials, no alerts
Uber's 10 November notice to the Iowa Attorney General names only "a service provider." It says that on 9 October the provider told Uber "an unauthorized third party had compromised the service provider's internal credentials and used them to access Uber's account with the service provider." Uber "immediately severed data connections with the service provider," and on 13 October its investigation found that the credentials had been used on 7 October to access Uber data connected to the provider. California's breach listing also gives 7 October; Uber's entry on Oregon's list gives 1 October as the breach date. Rescana's reading: 7 October is the last day of Fireworks' first incident window; neither Uber nor Fireworks has publicly linked Uber's exposure to either incident.
Uber says the data involved "images of driver's licenses and extracted text data, including name, address, date of birth, and driver's license number." The data sat in "a cloud storage environment created by Uber specifically to enable the service provider's work." "The environment had alerting enabled, but because the unauthorized third party used valid credentials and data flows enabled by the compromise on the service provider's system, no alerts were triggered." Uber says the credentials "did not provide access to other Uber systems."
Cursor: a key issued to Fireworks
Cursor's 13 November update says: "An attacker obtained an access key we had issued to Fireworks. The key was exposed by Fireworks, and the attacker used it to copy data from a read-only storage bucket we used to share data with them." The bucket held coding agent traces, mostly from 2025, from sessions not in Privacy Mode. Some traces held unredacted credentials (API credentials for "less than 0.01% of customers"), which the attacker used to reach select SaaS tools Cursor uses. Cursor deactivated the vendor's key, rotated credentials and moved all inference off Fireworks. It believes the incident is contained, with last attacker activity on 1 November, which was blocked.
Cursor's separate notice covers malicious versions 2.1.312 and 2.1.313 of the Anthropic.claude-code extension in its marketplace for Windows and macOS on 29 October, from about 04:27 to 15:35 EST (Rescana's conversion from Cursor's Pacific times). Its indicators include an svc-telemetry SSH key and a beacon on TCP 8099 to 134.209.82[.]225.
Notion: the timeline
Per Notion's advisory: on 2 November Fireworks told Notion a small set of data had been accessed, which Notion found was synthetic testing data. Notion disconnected from Fireworks on 3 November and deleted its data there on 4 November. On 6 November Fireworks said customer data may have been accessed, and on 8 November it confirmed the final dataset. Notion contacted impacted customers on 9 November, revoked exposed Notion API keys on 11 November, and on 12 November reported exposed third-party keys to their issuers. Notion used Fireworks for Insights classification and open-weight model inference; its FAQ says AI Meeting Notes data was not affected unless it was typed into a chat message or referenced in the AI's output.
Zero data retention: the vendor claim and the customer finding
Fireworks' 9 November entry says it has not detected impact to its Inference, Training SDK and Nexus products, "which are all zero data retention (ZDR) by default," and its 10 October entry says "Fireworks does not store prompts or outputs by default, as our platform is ZDR." Notion's Help Center FAQ tells a different story from the customer side. Notion says its own enterprise ZDR commitment does not cover Insights, where it retains requests and responses to deliver the feature and sent them to Fireworks for analysis, but that Fireworks "provided us with a contractual commitment to zero data retention for our use of its platform. Unfortunately, during our investigation we learned that data was retained from the Fireworks endpoints that we used." Notion is now auditing critical vendors, such as inference providers and data processors, "to verify that they abide by data minimization practices, including zero data retention policies."
Rescana's reading: Fireworks describes product defaults; Notion describes what actually held for its own endpoints, despite a contract. Neither has said publicly why the data was retained. Either way, a ZDR claim, default or contractual, is an assertion until you test it.
Affected Product Versions
We found no CVE and no CISA KEV entry for this incident. It is SaaS data exposure; the only versioned software involved is Anthropic.claude-code 2.1.312 and 2.1.313 in Cursor's marketplace. What each company lost is set out under Technical Information above.
Workaround and Mitigation
Fireworks customers that received a notice, per Fireworks' 10 October guidance:
- Replace revoked Fireworks API keys and update the applications that used them.
- Rotate named third-party credentials with their issuers. Fireworks warns: "Removing a stored secret alone does not invalidate the underlying credential."
- Review those providers' access logs, and rotate even if you find nothing.
- Check for unrecognized API keys, users, evaluators, training jobs and datasets.
Rescana guidance for every Fireworks customer. Given the 17 November scope, treat your account and billing data as exposed even without a notice, and expect phishing or invoice fraud that quotes real billing figures. Rotate every credential you stored as a Fireworks secret, named or not, since secrets were taken in both incidents and the second is still under forensic investigation. Then run the hunt below.
Notion customers, per Notion's guidance: no notice means Notion has not identified your workspace as affected. If affected, download the data from Security incident review in the workspace sidebar, treat it as exposed, and rotate any credentials found in it, highest-risk first. Notion says its annotated CSV's sensitivity flags are a starting point, and warns that names and email addresses from chats could be used in targeted phishing.
Uber drivers who received a notice can enroll in the free credit monitoring Uber offers. Cursor users who installed or updated the extension on 29 October should follow Cursor's notice; Cursor says admin rights are needed to enable SSH and that uninstalling alone may not be enough.
TPRM Lesson: AI Vendors Holding Sensitive Data
Our takeaways, drawn from the companies' statements:
- Inventory what each AI vendor holds, including data shared for training or offline analysis and secrets stored on the vendor's platform, not just live inference traffic.
- Scrub before sharing. Unredacted credentials in a small number of Cursor's traces opened its SaaS tools.
- Verify retention, not just the contract. See the ZDR contrast above.
- Scope every credential you issue to a vendor, and monitor its use for behavior, not just validity. Baseline vendor access (volume, timing, source) and alert on departures.
- Set notification deadlines and minimum content in the contract. Fireworks' first notice to Notion, on 2 November, said only that a small set of data had been accessed, which Notion's investigation found was synthetic test data; word of possible customer data came four days later.
- Have an exit ready. Cursor and Notion both moved inference off Fireworks.
Indicators of Compromise
Fireworks published an IOC PDF for customers on its security bulletins page. Fireworks says the indicators are "being continuously validated," so check the PDF for the current list. Below is a representative subset, defanged by Rescana; all indicators are credited to Fireworks unless marked.
Egress IP addresses (Fireworks lists 49 and recommends alerting or blocking), for example:
141.148.43[.]50,207.244.233[.]138,146.70.224[.]38,193.32.126[.]146,185.209.196[.]171,31.169.53[.]59,45.129.59[.]47,134.209.82[.]225
Shared Cloudflare egress (alert only; Fireworks warns these "may also carry legitimate traffic"), for example 104.28.243[.]187 and 104.28.225[.]223.
Outbound destination: 202.61.139[.]15 on TCP 8777, POST /exfil. Connections from your network to it are a potential signal.
User agents and behavior: firectl-admin/1.7.x and Mozilla/5.0 hunt; generic agents such as curl/8.14.1 and python-requests/2.34.2 are low fidelity on their own (Rescana's assessment; Fireworks notes that some IP addresses and user agents may represent shared infrastructure). Fireworks also lists API keys, users or service accounts you did not create, operator names containing "hunt," "probe" or "jump," one-off key validation calls (such as model listing) with no follow-up, evaluation jobs or evaluators you did not create, base64 blobs in logs under markers like *__DUMP_B64, and outbound requests to webhook[.]site.
Rescana's correlation: 134.209.82[.]225 is in Fireworks' egress list and is also the attacker endpoint in Cursor's extension notice. Neither vendor has publicly linked the two.
Where and how far back to hunt. Fireworks says to search your Fireworks account and the access logs of any third-party provider whose credentials you store as Fireworks secrets, as far back as retention allows. Rescana guidance: cover at least 3 August through 30 October 2026, the span of both incident windows, and keep provider logs under watch until the date you rotated each credential, since a stolen key works until it is revoked (Cursor saw attacker activity on 1 November). Fireworks advises treating a match as a trigger to investigate, not proof; treating no match as reassuring but not conclusive; and not deleting resources before you preserve log evidence. Report matches to Fireworks' security team at security@fireworks.ai with timestamps, resource IDs and source IPs.
What to Ask Your Suppliers Today
This is third- and fourth-party risk: your supplier's AI vendor held your supplier's data. Ask suppliers that use Notion or Cursor, or embed third-party AI models:
- Did you receive a notice from Notion, Cursor or Fireworks about incidents involving Fireworks AI? Date and named owner.
- Did you have a Fireworks account created before 30 October, and did you store our credentials as Fireworks secrets? If so, when were they rotated?
- Have you searched your Fireworks account and provider logs against Fireworks' IOCs from 3 August through at least 30 October? Result and date.
- Which AI inference providers and sub-processors handle our data? Is Fireworks AI among them, directly or indirectly?
- How do you verify ZDR? (Rescana guidance.) What testing, audit or contractual evidence shows a provider retains nothing, beyond its own statement?
- Would you detect a vendor using valid credentials in an unusual way?
- How quickly must AI sub-processors notify you, and how quickly do you notify us?
Track incomplete answers to closure. These are qualitative TPRM follow-ups, not a prevalence claim.
Need help working through SaaS suppliers and their AI sub-processors exposed to the Fireworks AI incident? Our team can help you prioritize them and track their answers. Talk to an expert
Sources
- Fireworks AI security bulletins: 17 November 2026 two-incident entry; 9 November and 10 October 2026 entries.
- Fireworks AI Indicators of Compromise (IOC) for Customers, PDF: indicators, hunt scope and handling advice.
- Uber notice to the Iowa Attorney General, 10 November 2026: provider notice, credential use, data, alerting, response.
- California Attorney General breach listing for Uber Technologies, Inc.: date of breach 7 October 2026.
- Oregon Department of Justice data breach list: Uber entry reported 10 November 2026, 149,346 affected, breach date 1 October.
- TechCrunch, 16 November 2026: Uber spokesperson names Fireworks.ai; no group has taken credit.
- Cursor, "An update on our recent security incident," 13 November 2026 and Cursor security notice: root cause, response, extension indicators.
- Notion security advisory: timeline and scope.
- Notion Help Center FAQ: review steps and the zero-data-retention finding.
- BleepingComputer, 11 November 2026 and TechCrunch, 12 November 2026: press coverage.




