OPAQUE, a confidential computing startup spun out of UC Berkeley, has joined two new Linux Foundation groups that want AI agent behavior to be independently verifiable. The company is now part of the Appia Foundation and the Agentic AI Foundation, known as the AAIF, the two bodies the Linux Foundation has built since December to referee how AI systems get assessed and how autonomous agents get governed. OPAQUE is contributing its own cryptographic proof specifications to both.
Google, Microsoft, OpenAI and Amazon Web Services already sit on the boards of one or both foundations. “AI is becoming part of the world’s critical infrastructure, and critical infrastructure can’t run on trust alone, it needs proof,” Aaron Fulkerson, OPAQUE’s chief executive, said in the announcement.
OPAQUE Puts Four Specifications Into Neutral Hands
OPAQUE’s contribution starts with the Agent Governance Toolkit, an open-source framework created by Imran Siddique, the company’s chief platform officer and formerly an agentic AI architect and engineering leader at Microsoft Azure. OPAQUE says the project has drawn close to 5,000 GitHub stars and more than 100 contributors.
Four specifications built on that toolkit are headed for the AAIF. The foundation already governs three founding projects, Anthropic’s Model Context Protocol (MCP), Block’s goose agent framework and OpenAI’s AGENTS.md convention, anchored by contributions the foundation announced in December.
- Agent Manifest gives every AI agent a verifiable identity.
- TRACE, short for Trust Runtime Attestation and Compliance Evidence, supplies hardware-attested trust records as cryptographic evidence of how an AI system ran.
- Confidential MCP, or cMCP, extends the Model Context Protocol with hardware-enforced policy and verifiable runtime evidence.
- Confidential A2A, or cA2A, verifies identity and provenance when AI agents communicate with each other.
Only TRACE has reached outside review so far, with key elements already contributed to CoSAI’s Working Group 4, a separate industry security effort. OPAQUE expects Agent Manifest, Confidential MCP and Confidential A2A to reach the AAIF in the coming months, once they mature into standards no single company controls.
Enterprises buying AI have no standardized way to verify how a model or agent actually behaved short of trusting the hyperscaler’s word for it, and no way to run models independently of the handful of providers currently gatekeeping access to them.
Fulkerson said that in the same announcement. He added that the foundation memberships give enterprises, regulators and auditors a way to check AI behavior without asking a vendor to vouch for itself.
Regulators Start Demanding Proof
The Appia Foundation launched a month before OPAQUE joined it, on June 17, hosted under the Linux Foundation’s Joint Development Foundation. Thirteen founding members signed on, including Arm, Ericsson, Google, Mastercard, Microsoft, Mitsubishi Electric, OpenAI, Schneider Electric and Siemens.
Its specifications provide testing criteria and component typologies for AI audits. They are split so an upstream provider’s conformity evidence passes down to whoever builds on top of it. No company needs to assess an entire system from scratch.
“As international standards and legal frameworks become more established, global organizations need a consistent, practical way to verify that AI systems conform to new expectations,” said Jim Zemlin, chief executive of the Linux Foundation.
Craig Shank, the Appia Foundation’s executive director, said, “AI systems now make decisions about people’s loans, their children’s schools and their jobs. People on the receiving end deserve to know those systems were built and assessed against criteria that hold up to scrutiny.”
Regulations are moving from principle to active enforcement, the Linux Foundation said when it announced Appia. Value-chain partners increasingly need evidence of trustworthy AI written into contracts and vendor reviews, not taken on faith.
Who Else Already Sits at the Table?
Both foundations already draw from a short list of the same technology giants. Google, Microsoft and OpenAI hold seats in the Appia Foundation and the AAIF at once, alongside sector specialists like Mastercard, Siemens, Anthropic, Block, Amazon Web Services, Bloomberg and Cloudflare, well before newer names like OPAQUE arrived.
| Detail | Appia Foundation | Agentic AI Foundation (AAIF) |
|---|---|---|
| Launched | June 17, 2026 | December 9, 2025 |
| Hosting structure | Joint Development Foundation, a Linux Foundation project | Directed fund under the Linux Foundation |
| Core focus | Conformity testing criteria and audit evidence for AI systems | Interoperable protocols for autonomous agents, including MCP, goose and AGENTS.md |
| Founding roster size | 13 members, including Arm, Google, Mastercard and Siemens | 8 platinum members, including AWS, Anthropic, Block and Bloomberg |
| Executive director | Craig Shank | Mazin Gilbert |
| Membership since launch | No growth figures reported yet | 190 organizations by May, over 200 by OPAQUE’s account |
The AAIF’s growth curve helps explain why OPAQUE wanted in early. Membership reached 146 organizations by late February, then climbed to 190 organizations by mid-May, adding gold members such as F5, GoDaddy, Stripe and TRON alongside financial and government names including JPMorgan Chase, American Express and the U.S. Army.
Financial firms have their own reasons to want independently checkable proof. Banks and payment networks, Mastercard on the Appia side, JPMorgan Chase and American Express within the AAIF, are already wrestling with a related problem: banks’ struggle to prove AI return on investment even when the technology performs as designed.
A Decade-Old Bet From a Berkeley Lab
OPAQUE’s underlying idea is nearly ten years old. Researchers at UC Berkeley’s RISELab built a system called Opaque in 2017 that ran analytics over data staying encrypted the entire time it was being processed, an approach known as oblivious computing.
The company built on that research is led by Ion Stoica, a UC Berkeley professor and co-founder of Databricks, Raluca Ada Popa, a senior staff research scientist at Google DeepMind who leads AGI security research, and chief technology officer Rishabh Poddar. Imran Siddique joined later as chief platform officer.
OPAQUE went to venture capital early and raised $31.6 million across three funding rounds, including a $9.5 million round in 2021 and a $22 million round in 2022, backed by investors such as Walden Catalyst Ventures and Intel Capital.
This summer, OPAQUE unveiled OPAQUE 3.0 on June 23 at the Confidential Computing Summit, an event the company created and now co-hosts with the Linux Foundation. The launch included Agent Manifest and Confidential MCP, with backing from the Technology Innovation Institute (TII), AMD and NVIDIA. Enterprises using the platform get a hardware-signed receipt any auditor can check without needing to trust OPAQUE itself, the company says. OPAQUE is also a member of the Confidential Computing Consortium.
Are Enterprises Actually Ready to Use It?
Not yet, based on independent research published this year. Multiple surveys found that most enterprises still lack the governance maturity to manage autonomous agents responsibly, even as they push AI systems into production, which is exactly the gap OPAQUE and the two foundations are aiming to close.
- 40% of enterprises will demote or decommission autonomous AI agents by 2027 because governance gaps only surface after something breaks in production, Gartner predicts.
- 65% of organizations say their current governance model was built for human decisions and does not translate to AI agents, according to an Omdia survey of 400 IT and data professionals.
- 66% of enterprises already allow AI agents into production without human review, or plan to within a year, while only 5% say they trust their own internal evaluations of those agents, a separate industry survey found.
Gartner’s research on AI agent governance failure blames a binary mindset. “Enterprises are treating AI agent governance as binary, either locked down or fully trusted, and that is the root cause of failure,” said Shiva Varma, a senior director analyst at Gartner.
The Standards Still Have to Graduate
None of OPAQUE’s four specifications are finished products. Only TRACE has reached outside review, through CoSAI’s Working Group 4. Agent Manifest, Confidential MCP and Confidential A2A remain inside OPAQUE’s own toolkit, expected to reach the AAIF in the coming months as the company and its partners mature them into standards nobody owns outright.
The risk is not new to open standards efforts. Vint Cerf, one of the internet’s original architects, raised a warning about fragmented AI agent protocols on his way out of Google, cautioning that the industry could repeat the early internet’s mistakes if agent standards multiply faster than anyone can govern them.
Imran Siddique framed OPAQUE’s wager plainly. “Verifiable trust in AI cannot be something any one company owns,” he said. “I’ve been contributing to both communities, and joining as a member is our commitment to build this in the open, with everyone, not around anyone.”
Whether the rest of the industry agrees in practice, rather than in a press release, will show up in how many of OPAQUE’s specifications actually reach the AAIF’s roster before the year is out.
Frequently Asked Questions
What Is the Agentic AI Foundation?
The Agentic AI Foundation, or AAIF, is a Linux Foundation body that launched in December 2025, co-founded by OpenAI, Anthropic and Block, to govern open protocols for autonomous AI agents. Its founding projects include Anthropic’s Model Context Protocol, which one industry analysis puts at more than 110 million monthly downloads by early 2026.
What Does the Appia Foundation Actually Assess?
The Appia Foundation organizes its specifications into two layers, a Requirements and Guidance layer and an Assessment Enablement layer, that translate international standards such as ISO/IEC frameworks into testing criteria auditors can apply to a specific AI model, system or application.
What Is Confidential Computing?
Confidential computing runs AI models and data inside hardware-isolated trusted execution environments, portions of a chip sealed off so that even the cloud provider hosting the workload cannot see the data or model while it runs, then produces a signed record proving the isolation held.
Can Any Company Join These Foundations?
Most can, for a fee scaled to size. Joining the AAIF requires Linux Foundation Silver Membership as a prerequisite, with dues based on a company’s total consolidated headcount, though associate membership is free for pre-approved nonprofits and universities.
What Is OPAQUE’s Agent Governance Toolkit?
The Agent Governance Toolkit, or AGT, is an open-source framework created by Imran Siddique that gives enterprise teams a common language for what an AI agent is allowed to do. OPAQUE describes it as a proving ground where new governance standards mature before graduating to bodies like the AAIF.
