Popular Topics
Chamberlain Hrdlicka Blawgs
Business and International Tax Developments Blog
What Fireworks AI and Optimizely teach us about managing risk in software procurement
On November 16, 2026, an attacker slipped malicious code into Optimizely’s Web Experimentation product. Because the code loaded from Optimizely’s own servers, browsers trusted it. A report found it running on the websites of 178 organizations where it skimmed payment card numbers and swapped cryptocurrency wallet addresses. None of those companies’ own systems were breached. Their vendor’s was.
A week earlier, AI inference provider Fireworks AI disclosed that an attacker reached its development machines on October 24 through a compromised third-party cloud provider. Notion, which uses Fireworks to run some of the models behind Notion AI, then disclosed that AI chat transcripts for under 1 percent of its users were exposed. Separately, Uber began notifying roughly 150,000 drivers that their names, dates of birth, and driver’s license numbers were stolen, and named Fireworks as the vendor involved.
The Chain Is Longer Than the Contract
Follow the Notion exposure. A Notion customer types into Notion AI. Notion sends the prompt to Fireworks. Fireworks runs development work on a cloud provider it has not yet named. The customer signed one contract. Its data crossed at least three companies.
That is now the normal architecture of software. Products are assembled from AI inference providers, analytics and testing scripts, content delivery networks, and cloud hosts, each with vendors of its own. Every layer adds risk the customer cannot see and its contract rarely reaches.
Uber said it had alerts in place, but none triggered because the attackers used Fireworks’ own internal credentials. Your monitoring cannot catch what looks like your vendor doing its job.
Lawyers Are Negotiating Products They Have Not Looked At
In practice, the biggest gap in procurement is rarely a missing clause. It is a disconnect between the lawyers and the technology, on both sides of the table. Buy-side counsel negotiates a data processing agreement and a security exhibit from a template without knowing whether the product sends customer data to a third-party AI model or injects code into the customer’s website at runtime. Sell-side counsel signs representations about subprocessors and security controls that the engineering team was never asked to confirm.
The result is paper that looks complete and protects very little. An Optimizely customer that never asked how vendor scripts are delivered had no reason to negotiate controls over them. A Notion customer that never asked where AI prompts go had no way to know Fireworks held its transcripts.
What Businesses Should Do Now
- Map the data flow before you negotiate.
- Put a technical reviewer in the room.
- Make flow-down obligations real.
- Limit retention downstream.
- Control what third-party code can do on your site.
- Update legal review triggers.
- Review their own insurance coverage with the expectation that consolidated vendors may be underinsured themselves.
Vendor risk used to be a question about one company. It is now a question about a supply chain. The contracts need to catch up.
- Senior Counsel
Marcus Burnside advises technology companies, private equity-backed businesses, and foreign clients on intellectual property strategy, AI governance, and data privacy. His practice sits at the intersection of three areas most ...



