Braintrust AWS account compromise and customer API key rotation

Summary

On May 5, 2026, Braintrust, an AI evaluation and observability platform, posted a website notice disclosing unauthorized access to one of its AWS accounts that held customer API keys for cloud-based model providers. The company emailed every customer the next day asking them to rotate any keys stored with the platform. Braintrust told TechCrunch it had identified one directly impacted customer and that the wider notice was sent out of caution while the investigation continues.

What happened

• Braintrust observed suspicious activity in one of its AWS cloud accounts on May 4, 2026, and confirmed unauthorized access shortly afterward.
• The affected account held customer API keys that the platform used to call third-party AI model providers on customers' behalf.
• Braintrust said it locked down the account, audited and restricted access on related systems, and rotated its internal secrets.
• On May 6, 2026, the company emailed all customers asking them to rotate any keys stored with the platform.

Timeline

• 2026-05-04 - Braintrust observes suspicious activity in an AWS account and begins investigation.
• 2026-05-05 - Braintrust posts a website notification confirming an incident.
• 2026-05-06 - Braintrust emails customers requesting key rotation; TechCrunch and other outlets publish accounts of the disclosure.

What the vendor has confirmed

Braintrust spokesperson Martin Bergman told TechCrunch the company had "confirmed a security incident" and that the customer notice was sent "out of an abundance of caution." Braintrust said the incident had been contained and the cause was still under investigation. The company stated that as of disclosure it had identified one directly impacted customer and had no evidence of broader exposure.

What remains unclear

• The root cause and the duration of the unauthorized access have not been disclosed.
• Braintrust has not published the count of customers whose keys were stored in the affected account.
• Whether the exposed keys were used against the third-party model providers they unlocked has not been disclosed.

Broader context

API keys held by AI middleware platforms fan out to every model and tooling vendor those keys unlock, so a single account compromise at the orchestrator can hand an attacker downstream access to multiple services without separately reaching each one. This concentration of credentials in evaluation, monitoring, and routing layers has been a recurring concern as the AI tooling stack has grown.

Sources

• TechCrunch coverage with vendor statement (press)
• Prism News coverage (press)