cd /news/ai-safety/the-fireworks-ai-breach-how-one-vend… · home › topics › ai-safety › article
[ARTICLE · art-172764] src=blog.gitguardian.com ↗ pub= topic=ai-safety verified=true sentiment=↓ negative

The Fireworks AI Breach: How One Vendor Key Became a Supply Chain Attack on Cursor

A credential chain that began with the October 10, 2026 Fireworks AI breach led to trojanized Claude Code builds sitting in Cursor's marketplace cache for about 11 hours on October 29, 2026 — between approximately 2:27 a.m. and 1:35 p.m. Pacific time — potentially reaching up to 5,000 developers, according to GitGuardian's account of the incident. Cursor said a read-only storage key it issued to Fireworks was used to copy coding agent traces, and unredacted credentials inside those traces exposed API credentials for fewer than 0.01% of Cursor customers and gave the attacker access to Cursor's SaaS tools. Fireworks confirmed by November 17 that account and billing data were accessed for all accounts created before October 30, while Notion separately reported that chat transcripts from fewer than 1% of workspaces were exposed and revoked the affected Notion API keys.

by read10 min views3 publishedNov 19, 2026
The Fireworks AI Breach: How One Vendor Key Became a Supply Chain Attack on Cursor
Image: Blog (auto-discovered)

TL;DR #

  • One read-only key started the chain: after Fireworks was breached, the attacker used a key provided by Cursor to copy agent traces. The credentials found in those traces gave access to Cursor’s internal tools.
  • Up to 5,000 developers may have received a malicious extension: trojanized Claude Code builds sat in Cursor’s marketplace cache for about 11 hours on Oct. 29. The malicious builds were available between approximately 2:27 a.m. and 1:35 p.m. Pacific time on October 29, 2026.
  • Notion’s exposure was separate: chat transcripts from fewer than 1% of workspaces were exposed. Notion revoked exposed Notion API keys and notified the issuing providers about exposed third-party API keys.
  • Where GitGuardian fits: Developer Endpoint Protection shows which credentials sit on developer machines, and AI Hooks keep new secrets out of supported coding assistants.

On October 10, Fireworks AI disclosed unauthorized access to its systems. Over the following weeks, Cursor and Notion disclosed how the breach affected them, and Cursor took down a trojanized Claude Code extension from its marketplace. The attacker moved from Fireworks into Cursor’s systems by using credentials stored in places their owners thought were secure.

Below is a step-by-step overview of the credential chain, what the malicious extension did, and recommended changes for sharing credentials with AI vendors and coding tools.

What happened in the Fireworks AI breach #

Fireworks AI, an inference and fine-tuning provider, has reported two separate incidents.

The scope expanded with each update. The initial notice stated that environment variables, including API keys and tokens stored as secrets, were captured from a limited number of accounts. By Nov. 17, Fireworks confirmed that account and billing data were accessed for all accounts created before Oct. 30, and that datasets, models, or secrets were accessed for a small number of customers.

Each disclosure points to the same weak spot: credentials and data held by a vendor. Fireworks says its own credentials were used to reach customer secrets, Cursor says a key it issued was used to copy its agent traces, and Notion says chat transcripts containing pasted keys were accessed.

How the attacker got from Fireworks to Cursor’s marketplace #

The attacker gained initial access to Fireworks. Then, they used valid credentials wherever they were accepted.

The attacker reached secrets stored at Fireworks

Fireworks customers stored API keys and third-party tokens as secrets so evaluation jobs could call other services. When the attacker accessed these jobs, the secrets were also compromised. Fireworks advises that deleting a stored secret from its platform does not invalidate the credential itself. The credential remains active at the issuing provider until it is rotated there.

Cursor’s read-only key exposed a bucket of agent traces

Cursor issued Fireworks an access key to a read-only storage bucket used for data sharing. After obtaining the exposed key, the attacker used it to copy the bucket's contents. Read-only access was sufficient, as the value was in the data stored within the bucket.

Credentials in those traces opened Cursor’s SaaS tools

The bucket contained coding agent traces, primarily from 2025 sessions outside Cursor's Privacy Mode. Cursor said these sessions were eligible for use in model training.

Some traces included unredacted credentials, resulting in fewer than 0.01% of Cursor customers having API credentials exposed. The attacker used these credentials to access SaaS tools used by Cursor, then archived source code, support conversations, and credentials users provided to Cursor support.

This scenario is familiar to engineering teams. A developer may paste a token into a prompt, or an agent may read a .env file during debugging. The secret then appears in a session transcript, which can be stored, shared with a model provider, or used for training.

A credential stolen from Cursor enabled the marketplace compromise

In the last step, the attacker used a credential taken from Cursor’s internal data to put trojanized Claude Code builds on the cache that serves Cursor’s extension marketplace. Developers who installed or updated the extension from that marketplace during an 11-hour window on Oct. 29 could have received these compromised builds.

What the malicious extension did, and what it didn’t #

Cursor and Microsoft have identified the scope of the malicious builds, which affect versions 2.1.312 and 2.1.313 for both macOS and Windows:

  • Potential impact: Cursor reports that the incident could have affected as many as 5,000 users, but this is the maximum estimate and not a confirmed number of compromises.
  • Malicious activity: Microsoft found that the malicious builds tried to add an attacker’s SSH key, turn on the local SSH service, set up reverse SSH forwarding, create persistence, and send host details. Cursor states that activating the SSH service and completing remote-access steps would fail on most machines because they require administrative privileges. However, the extension can still write SSH keys, create persistence artifacts, and transmit host information.
  • Unaffected sources: The malicious builds were not shared through Microsoft’s Visual Studio Marketplace or the Claude Code npm package. Anthropic did not take part in their development.
  • Persistence risk: Cursor warns that just uninstalling the extension might not be enough, since the payload installs separate scripts that keep running. Theyrecommend changing credentials and sessions that could be reached from affected machines.

What about Notion? #

Notion used Fireworks for classification tasks supporting the Insights feature and as an inference provider for open-weight models. This exposure was unrelated to the Cursor attack chain. Although Fireworks held the affected data, there is no evidence linking the Cursor compromise to the Notion exposure.

  • What was exposed: Exposed data included Notion AI chat transcripts used for the Insights feature, impacting fewer than 1% of workspaces. Most transcripts were not attributable to specific users or workspaces and did not contain tool-call results.
  • Why it matters: Some affected messages included passwords, API keys, or other credentials entered by users into Notion AI.
  • What Notion did: Notion revoked exposed Notion API keys found in affected messages on November 11 and reported exposed third-party API keys to their issuing vendors on November 12.

Notion disabled Insights during the investigation and stopped serving open-weight models through Fireworks. The company recommends that affected customers rotate any credentials entered in the impacted conversations.

What AppSec teams should change after the Fireworks breach #

You cannot patch a vendor's infrastructure, but you can control which credentials you share, how quickly you can revoke them, and how many are exposed in AI sessions.

Detect: list what you’ve shared

Start by listing all credentials your organization has issued to AI providers, not just those received from vendors. For example, Cursor’s exposure began with a key it provided to Fireworks. Include model API keys, storage keys for shared datasets, and third-party tokens stored in the provider’s secrets manager. Specify the access level for each credential.

When a vendor discloses an incident, identify which credentials were accessible through the affected systems and prioritize revocation based on exposure, permissions, and potential impact. Read-only doesn’t mean low-risk. A read-only key to a bucket of agent traces was enough here.

Next, assess the developer environment. Coding agents may write credentials into transcripts, local configurations, and shell history that are not detected by repository scanners.

If a developer installed or updated Claude Code versions 2.1.312 or 2.1.313 from Cursor's extension marketplace on macOS or Windows between 2:27 a.m. and 1:35 p.m. PT on Oct. 29, 2026, follow Cursor's remediation instructions:

  1. Preserve the extension package, modified SSH files, generated keys, and scripts before cleanup.
  2. Isolate machines that show evidence the payload ran or the backdoor was installed.
  3. Remove the extension, the unauthorized SSH key, the persistence mechanisms, and any running tunnel processes.
  4. Revoke credentials and sessions accessible from the machine, from a trusted device.
  5. Rebuild the machine if the investigation can't rule out persistence.

Remediate: rotate where the key was issued

Removing a secret from the vendor's platform does not affect the credential itself. Rotate the credential with the issuing provider, then review that provider's logs for reads, downloads, and new resources. Fireworks discovered that the attacker had created new API keys, so revoking the original key may not eliminate all unauthorized access. Notion revoked exposed Notion API keys and reported third-party keys to the respective vendors.

Speed is critical because leaked keys can remain valid for years. Per GitGuardian's State of Secrets Sprawl 2026, of the secrets GitGuardian confirmed as valid in 2022, 64% were still unrevoked in January 2026. The main challenge is not rotating the credential, but identifying every copy and owner.

Prevent: limit what vendors hold

Share only the minimum necessary key, limit data-sharing buckets to required access, and use Privacy Mode or zero-retention settings when available. Cursor's exposed traces originated from sessions outside Privacy Mode.

Data shared in the past remains vulnerable to theft. Set retention limits and scan existing datasets for secrets before they leave your environment.

How GitGuardian can help #

The attacker didn't need a new exploit at every stage. Credentials exposed in one system provided access to the next. That's where stronger credential security can make a difference: reducing the secrets available to attackers, finding exposures across systems, and remediating them before they enable further compromise.

  • Prevention in AI workflows:GitGuardian AI Hooks can block secrets before they enter prompts and tool calls in supported coding assistants, reducing the risk of credentials appearing in vendor-stored traces. Blocking capabilitiesdiffer by assistant , as each exposes unique hook points.
  • Endpoint visibility: Developer Endpoint Protection shows which credentials are stored on a developer’s machine, including local configs, shell history, IDE settings, and agent files. If a laptop is compromised by a malicious extension, these credentials should be prioritized for revocation.
  • Detection beyond code: Internal Secrets Monitoring extends to collaboration tools such as Slack, Jira, and Confluence. In the Cursor compromise, credentials shared through support conversations created another source of exposure, illustrating why secrets detection needs to extend beyond code repositories.
  • Remediation: Identifying exposed credentials is only the first step. Each credential requires an assigned owner and a clear process for revocation with the issuing provider.

Credentials connect your vendors, so breaches can spread through those links. In the Fireworks breach, a key Cursor had issued to Fireworks led the attacker from one credential to the next.

Those credentials are part of your credential layer: the credentials that let code, machines, developers, and AI agents access systems, wherever those credentials live or move.

To stop a breach from spreading, identify which credentials link your systems to others and revoke them at the source.

FAQ #

What happened in the Fireworks AI breach? #

Fireworks AI experienced two security incidents in 2026. From Aug. 3 to Oct. 7, attackers used stolen credentials and exploited a service flaw to collect environment variables, including API keys and tokens. Beginning Oct. 23, a breach at a third-party cloud provider allowed attackers to access Fireworks development machines. Account and billing data for all accounts created before Oct. 30 were also compromised.

Was Cursor hacked, and what did attackers get? #

Cursor reported that an access key provided to Fireworks was exposed by Fireworks. The attacker used this key to copy coding agent traces, some of which contained unredacted credentials. The attacker also accessed archived source code, support data, and a credential that enabled them to place a trojanized Claude Code extension in the marketplace cache for approximately 11 hours. Cursor stated there was no impact on Privacy Mode data, customer code repositories, or its production systems.

Was Notion affected by the Fireworks AI breach? #

Yes. Notion reported that chat transcripts from fewer than 1% of its workspaces were exposed. Notion had used Fireworks as an inference provider for open-weight models. It revoked exposed Notion API keys found in affected messages and reported exposed third-party API keys to the vendors that issued them.

What is LLMjacking and how do attackers monetize stolen AI keys? #

LLMjacking refers to stealing credentials that allow access to large language models. Attackers may use these stolen credentials themselves or sell them through gateways that provide inexpensive access to the models, resulting in the victim incurring the cloud charges.

── more in #ai-safety 4 stories · sorted by recency
── more on @fireworks ai 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/the-fireworks-ai-bre…] indexed:0 read:10min 2026-11-19 · —