Anthropic Absorbs Opaque Systems, the Confidential-Computing Startup Behind TRACE and Confidential MCP

Anthropic has quietly taken in Opaque Systems, a confidential-computing startup whose technology lets AI systems work on sensitive data while the underlying hardware keeps that data shielded. SecurityWeek reported the transaction on October 19, noting that financial terms were not disclosed and that Opaque itself announced the move about a week earlier: its team is joining Anthropic, and Opaque Systems is winding down its operations.

The startup carries a decade of academic lineage. It began as a research project at UC Berkeley's RISELab and was co-founded by Ion Stoica, who also co-founded Databricks and Anyscale, and Raluca Ada Popa, an encryption researcher who now works at Google DeepMind. Enterprise veteran Aaron Fulkerson, previously a ServiceNow executive, took over as chief executive in October 2023, with Rishabh Poddar as CTO. Earlier this year Opaque raised a $24 million Series B at a $300 million valuation, and its customer list included Anthropic itself, alongside ServiceNow and Accenture.

Opaque started with confidential computing, using secure enclaves on chips from Nvidia and Intel and working with Microsoft and Google on the surrounding ecosystem. As autonomous agents spread, the company repositioned around what it called verifiable AI: replacing trust assumptions with cryptographic evidence of what ran, where it ran, and under which policies. Its own wind-down statement put the pivot bluntly - protecting data and models was no longer enough, because organizations now need to verify how agents behaved, not just where the data sat.

What happens to the open-source work is the part developers care about. Opaque says its projects continue at AgenTrust under their existing licenses: Confidential MCP, which runs Model Context Protocol tool calls inside attested enclaves; Confidential A2A; TRACE, an open standard for AI runtime attestation; and the Agent Manifest and Weight Custody Manifest specifications. The Confidential Computing Summit continues under the Linux Foundation.

Reporting on the structure differs in degree rather than substance. Constellation Research described the arrangement on October 13 as a technology license plus a hire of key staff, and The New Stack called it an acquihire; both agree that Fulkerson, Poddar and a group of engineers move to Anthropic while the corporate entity does not, and that not everyone at Opaque is making the trip. Neither party has said what the wind-down means for existing customer contracts, which leaves Opaque customers needing a migration plan.

The strategic logic is easy to read. Banks, hospitals and governments want frontier models on data they cannot legally hand over in plain form, and proof of where a model ran and what it touched is becoming a procurement requirement for sovereign and on-premises deployments. The same applies to agents: as they gain access to email, codebases and payment systems, buyers want attested, auditable runs rather than logs they have to take on faith. Anthropic has been pushing hard into enterprise and government accounts, and absorbing the attestation layer in-house signals it expects verifiability to be a product feature, not a third-party add-on.

Two threads are worth watching. One is whether Anthropic folds the attestation work into Claude for enterprise and government offerings, where Opaque's technology would map directly onto existing compliance demands. The other is whether TRACE and the other open-source projects keep receiving real maintenance under AgenTrust, or drift quietly now that the company behind them is gone. For teams running agents in regulated sectors, both answers matter more than the undisclosed price.