Fireworks AI: Cloud Breach Exposes API Keys

At-A-Glance Incident Diagnostic Telemetry

---

Executive Incident Briefing (2-Minute CISO Brief)

1. Strategic Incident Overview

Fireworks AI, a cloud-based AI infrastructure startup, has disclosed a security breach in which attackers accessed its development machines after breaching a third-party cloud provider. The breach exposed API keys, tokens, and other sensitive data, including resource metadata, billing information, and incomplete model weights and datasets for some customers.

The company has notified law enforcement and customers, and has taken remediation measures to close known critical vulnerabilities and restrict unauthorized network pathways. Fireworks is working with Palo Alto Networks' Unit 42 incident response division to secure its systems against future attacks.

The breach is attributed to a financially motivated cybercriminal threat actor group, which is aimed at selling tokens from unauthorized access to Large Language Models (LLMs) to illicit routing gateways, known as 'LLMJacking'. The incident highlights the risks of cloud-based AI infrastructure and the importance of robust security measures to protect sensitive data.

2. Business Impact & Perimeter Exposure

The breach has significant business implications for Fireworks AI, including potential reputational damage and financial losses. The company has taken remediation measures to close known critical vulnerabilities and restrict unauthorized network pathways, but the incident may still have a lasting impact on customer trust and confidence.

3. Compliance & Governance Blast Radius

The breach may trigger applicable regulatory requirements, such as GDPR Article 33, which mandates notification of personal data breaches to the relevant supervisory authority. Fireworks AI may also be subject to other regulatory requirements, such as the California Consumer Privacy Act (CCPA) or the Payment Card Industry Data Security Standard (PCI DSS).

---

Breach Architecture & Blast Radius Analysis

Compromise Vector & Exposure Mechanics

The breach occurred when a third-party cloud provider that hosted Fireworks R&D machines experienced a security incident on October 23. The unauthorized third party used this access to reach Fireworks development machines and run an unauthorized workload, leading to additional customer data impact.

The attackers accessed environment variables containing API keys and tokens stored as secrets for a limited number of accounts. They also accessed resource metadata, billing information, and incomplete model weights and datasets for some customers. However, the company said the attackers couldn't access customer payment card numbers because Fireworks doesn't store them.

The incident highlights the importance of secure storage and management of sensitive data, such as API keys and tokens. It also underscores the need for robust security measures, including incident response processes and vulnerability management, to protect against unauthorized access and data breaches.

According to researchers at Google's Threat Intelligence Group (GTIG) and Mandiant, large language model hijacking attacks (also known as LLMjacking) are driving record costs for compromised organizations, with cloud AI charges reaching as high as $1 million due to fraudulent calls using stolen API keys.

Adversaries frequently leverage compromised credentials, infostealer session tokens, or misconfigured cloud access policies to bypass perimeter defenses. Once inside cloud tenancies, threat actors execute automated enumeration to exfiltrate database records and object storage repositories.

Impacted Asset Telemetry & Exfiltration Scope

Forensic analysis indicates targeting of customer identity tables, corporate emails, and API configurations, exposing affected organizations to downstream credential-stuffing and targeted spear-phishing campaigns.

---

MITRE ATT&CK Tactics, Techniques & Procedures (TTPs)

---

Threat Actor Attribution & Extortion Tactics

The breach is attributed to a financially motivated cybercriminal threat actor group, which is aimed at selling tokens from unauthorized access to Large Language Models (LLMs) to illicit routing gateways, known as 'LLMJacking'. The group is likely to be sophisticated and well-resourced, given the complexity of the attack and the scope of the data breach.

---

Incident Response & Containment Playbook

1. Credential Invalidation & Session Termination

• Force tenant-wide credential resets and invalidate all active session tokens and OAuth refresh secrets.
• Revoke all long-lived API keys, IAM access tokens, and administrative credentials across cloud environments.

2. Cloud IAM & Storage Bucket Hardening

• Enforce strict least-privilege role-based access control (RBAC) and disable public access on all cloud storage buckets.
• Require conditional access policies requiring managed, compliant devices for administrative console access.

3. Compromise Verification & Identity Hunting

Microsoft Sentinel Hunting Query (KQL): Anomalous Bulk Download

CloudAppEvents
| where ActionType in~ ("FileDownloaded", "DataExported", "StorageBlobRead")
| summarize DownloadCount = count() by AccountName, IPAddress, bin(TimeGenerated, 1h)
| where DownloadCount > 100
| order by DownloadCount desc

4. Verification & Audit Checklist

• Review and update incident response processes to ensure timely detection and response to security incidents
• Conduct regular vulnerability assessments and penetration testing to identify and address potential security weaknesses
• Implement robust security measures, including multi-factor authentication and encryption, to protect sensitive data
• Review and update access controls and permissions to ensure that only authorized personnel have access to sensitive data
• Consider implementing a bug bounty program to encourage responsible disclosure of security vulnerabilities
• Review and update cloud security configurations to ensure that they are aligned with industry best practices
• Consider implementing a cloud security gateway to monitor and control cloud-based traffic

---

Exposure Telemetry & Breach Indicators

Review tenant audit logs to identify anomalous data egress originating from unmanaged IP addresses.

---

Frequently Asked Questions (FAQ)

1. Was sensitive customer or corporate data exfiltrated during the Fireworks AI breach?

Forensic investigations confirm unauthorized access to corporate cloud repositories and database records, creating immediate risks of credential-stuffing and targeted spear-phishing campaigns.

2. What initial compromise vectors were leveraged by the threat actors?

Threat actors commonly exploit compromised identity tokens, unrotated service account credentials, or misconfigured cloud storage access policies to achieve unauthorized cloud tenant ingress.

3. What immediate containment steps should security teams execute?

Immediately invalidate all active session tokens, rotate tenant-wide administrative credentials, enforce least-privilege IAM policies, and restrict cloud storage buckets to internal VPC endpoints.

4. What regulatory disclosure requirements apply to this incident?

Public enterprises must evaluate materiality under SEC Item 1.05 of Form 8-K (4 business days), while GDPR Article 33 mandates supervisory authority notification within 72 hours if personal data is affected.