Cursor Breach: Old AI Traces Exposed Live Credentials
Cursor links its breach to credentials in historical AI traces. Separate retained copies, API-key revocation and malicious-extension cleanup.
Cursor Breach: Old AI Traces Exposed Live Credentials
Cursor says a key exposed at Fireworks let an attacker copy historical coding-agent traces containing credentials, then reach other systems. The lesson is not that every Cursor conversation leaked: it is that an old AI record can still contain usable authority. Treat exposed secrets differently from ordinary text, and assess malicious-extension exposure separately.
Cursor published its fuller investigation on November 13, 2026. This guide is for people who use AI with work files, developer tools or support logs—and anyone deciding how much context to put into a hosted assistant.
We reviewed the published incident documents, not affected accounts or devices. The original worksheet below helps separate the response tasks without claiming we independently verified the companies' forensic findings.
What Is Confirmed
Cursor's November 13 report describes a read-only bucket holding limited traces, mostly from 2025, from sessions outside Privacy Mode. Unredacted credentials led into SaaS tools and then to support/business records and archived non-production code. It reports API-credential exposure for fewer than 0.01% of customers and trace exposure for fewer than 0.05% of non-Privacy-Mode users.
The company says Privacy Mode prompts/code, active inference, customer repositories and official app updates were unaffected. It says the last observed attacker activity, November 1, was blocked and believes the incident contained. These are company findings, not an independent guarantee.
Separately, the original extension notice identifies tampered Anthropic.claude-code versions 2.1.312 and 2.1.313, delivered through Cursor's VS Code extension marketplace on October 29, approximately 2:27 a.m.–1:35 p.m. Pacific, on Windows and macOS. The unauthorized builds attempted a persistent SSH backdoor. Cursor says these were not developed by Anthropic and does not identify the official distribution as affected.
Do not apply either percentage to every kind of exposure. A trace-related percentage is not the rate of extension installation, successful backdoors or compromised developer machines.
What Is Still Unclear
Public reporting does not let us determine whether a particular reader's secret was copied, whether a copied credential was later used, or whether an individual machine executed a payload. Those require account-specific notices and relevant logs.
Potential delivery, execution, persistence and successful remote access are different outcomes. A count for one is not a count for the others. A privacy setting for trace storage does not answer whether a separately distributed extension ran.
We also cannot infer that changing a privacy setting today retroactively removes an older exported dataset. That would need a documented deletion process and verification of the particular retained copies.
The important contrary evidence is already in the scope above: this is not a finding that all live requests, Privacy Mode content, official releases or customer repositories were breached.
Why a Read-Only Archive Can Still Carry Authority
Read-only storage constrains changes to that storage. It does not make every readable value harmless.
Consider a hypothetical support transcript containing a still-valid deployment token. A reader who cannot alter the transcript might nevertheless submit that token to the service it unlocks. The storage permission and the credential's permission are separate boundaries.
That gives AI-workflow reviews two questions:
- What information is in the copy? Personal details, proprietary text, prompts, tool output, file fragments or troubleshooting context can all matter.
- What authority is in the copy? Tokens, session material and credentials can authorize actions beyond merely revealing the conversation.
The distinction applies whether the copy is called a trace, evaluation sample, attachment or support record. Renaming it does not change what its contents permit.
This is analysis of a general failure pattern, not a claim that every AI trace contains credentials or that every exposed key remains valid.
Use Three Separate Response Tracks
Do not reduce a combined incident to a single action such as deleting a conversation or uninstalling an extension. Use a record of three independently verifiable tasks:
| Track | Question to answer | Evidence to keep | What does not close it |
|---|---|---|---|
| Retained copy | Which dataset, support record or export contained your information? | Notification, relevant dates, data categories and provider response | Changing a future collection preference alone |
| Credential authority | Which issuing service still accepts the exposed secret? | Revocation/rotation result and a review of use during the exposure window | Deleting the text that once displayed it |
| Endpoint persistence | Did a malicious package execute or leave independent artifacts? | Installation history, security findings and authorized remediation results | Removing the visible extension alone |
This is our response worksheet, not a vendor dashboard or an automated incident test. One track can be complete while another remains open.
For example, an administrator might revoke an affected API credential but still need to investigate the workstation that held it. Conversely, a clean workstation inspection does not decide what happened to an older provider-side dataset.
What Affected Readers Should Do
Use the current vendor notice and your organization's security process, rather than improvising destructive cleanup from a blog post.
Cursor recommends checking telemetry, isolating affected hosts, removing unauthorized keys and persistence, assessing rebuilding, and rotating accessible credentials from a trusted system. Uninstalling alone may leave standalone scripts behind.
Cursor reports coding errors and missing administrative rights prevented SSH enabling on many hosts, but still calls for remediation. A failed tunnel is not a reason to skip assessment.
For teams, assign the three tracks to their actual owners. The person managing an AI subscription may not control your cloud keys or endpoint security. Record who verifies each outcome, and avoid putting replacement secrets into the incident conversation itself.
Do not run a cleanup script simply because someone reposted it. Obtain current instructions from the official notice, have the appropriate administrator review them, and preserve evidence your incident-response process requires.
Training Archives and Inference Are Different Data Paths
Fireworks' November 9 update describes impacts involving account/billing information and, for limited accounts, secrets, datasets and incomplete models. It says it had not detected impact to Inference, Training SDK or Nexus, which it describes as zero-data-retention by default, while its investigation continued.
Those distinctions matter. A statement about a live inference product is not automatically a statement about every historical dataset, evaluation store, support system or separately shared archive.
When comparing AI tools, ask about the object and its purpose rather than stopping at a label:
- Is this the active request, saved conversation, model-improvement dataset or support attachment?
- Who can access each copy, and what credentials grant that access?
- Does the setting govern future collection, existing copies, or both?
- Which deletion or retention process applies to each object?
- If the object contains a secret, who can revoke its authority independently of deleting it?
For a related but distinct example, our Notion/Fireworks transcript guide explains why missing account labels do not necessarily make conversation content anonymous. Do not merge the affected populations of these two incidents.
Where OpenVeil Fits—and Where It Does Not
If your task is brainstorming, drafting or comparing information rather than operating a developer workstation, OpenVeil offers adults a hosted, privacy-focused conversational workspace.
Its documented boundary is specific: private chat history stays in your browser rather than becoming a server-side chat-history record. OpenVeil does not use prompts, uploads or outputs to train foundation models. Active requests still require processing by OpenVeil and the providers needed to fulfill them; account and billing records are separate.
That narrower history policy is relevant when choosing how to keep working context. It is not a reason to paste credentials into OpenVeil or any other assistant. Browser-local history does not make hosted processing fully offline or anonymous, and clearing browser data or switching browsers can affect access to that history.
OpenVeil is not endpoint-security software, a credential vault, a Cursor patch or a remedy for an already exposed archive. Read what browser-local history means and what to check before trusting an AI privacy claim. If that workflow fits, try OpenVeil with a non-sensitive draft—not a live access key.
Frequently Asked Questions
Does deleting an AI chat revoke a leaked API key?
No. Removing a record and invalidating a credential are different operations. Verify revocation with the service that issued the credential and review relevant access records.
Does read-only storage mean exposed material cannot be used elsewhere?
No. Read-only describes permission on that storage location, not the permissions carried by a credential inside it.
Is a clean extension list enough to close a malware investigation?
Not necessarily. An installed package can create artifacts outside its own directory. Follow the applicable advisory and have the endpoint owner verify the result.
Does this prove that private AI chat is impossible?
No. It supports asking precise questions about copies, purposes and authority. Different storage and processing designs have different boundaries; none should be reduced to a slogan.
Bottom Line
An old AI record can hold more than information: it can hold a credential whose authority must be revoked separately. Review retained copies, credential permissions and endpoint artifacts as distinct tasks. Do not mistake one successful cleanup step for completion of all three.
Try OpenVeil Plus for $10/month
If the broader issue is choosing an everyday AI workspace without a normal cloud-chat archive, OpenVeil provides hosted chat with browser-local private-chat history and no normal server-side chat-history record. It does not patch another product, remove malware, secure a compromised device, or guarantee immunity to prompt injection. Complete the security actions in this article first.
OpenVeil Plus is $10/month with 10,000 monthly credits and the same feature set as the higher paid plans. Usage consumes credits; this is not unlimited access.
Once your environment is safe, try a non-sensitive question rather than uploading credentials or private incident evidence. Create your OpenVeil account and try the free preview: ten bounded actions, no card required; video is paid-only. Ready to continue? Sign up for OpenVeil Plus at $10/month, then choose Plus in Credits & billing to complete checkout.
OpenVeil is for adults 18+. Active requests still require OpenVeil and necessary provider processing. Read the privacy policy before using sensitive material.
Sources
- Cursor: November 13 investigation update.
- Cursor: October 29 extension security notice and current remediation guidance.
- Fireworks AI: November 9 incident update.
Research cutoff: November 15, 2026. Follow current official notices for account-specific scope and remediation; this article is not a device assessment or proof of credential misuse.