Upcoming changes in the development process of our teams working on software for the Tezos L1 blockchain and Etherlink

This is a joint post from Nomadic Labs, Trilitech and Functori.

TL;DR: AI agents have made finding and exploiting software vulnerabilities faster and cheaper than ever, across the whole software industry, not just blockchains. The same tools have also opened up new ways to check for bugs and defend against attacks, and we have leaned into this heavily. Over 2026, we have used AI to build a continuous audit process for our work on the Tezos L1 protocol and Etherlink, which has found and fixed several issues before they became a problem. But the rapid growth in the use of AI agents also raises the risk that public development branches and discussions may expose security flaws before fixes are deployed. To reduce that risk, as part of a broader strategy to evolve our engineering processes, we are moving our day-to-day development on Tezos L1 and Etherlink from their current upstream location to a private repository starting December 2, 2026. Released source code will remain publicly available under the MIT license, and we are committed to improving the quality of any tools and documentation that we create to help enable anyone to submit protocol proposals through Tezos governance.


2026 has marked a turning point in how software gets built, for us and the rest of the industry. A few core engineers at Nomadic Labs, Trilitech and Functori had already been experimenting with tools like Claude Code and Codex during 2025. They had seen how transformative they could be with the right approach, and witnessed the rapid progress of the frontier models. Over the Christmas holidays, experimentation spread across our teams and by the time we returned to work in January, AI agents were no longer niche.

Like everyone else, we had to figure out what that shift meant in practice. These tools have changed not only how we write code, but also how we review it, test it, and find bugs in it. One thing became very clear: their impact goes well beyond producing new software, and cybersecurity is the clearest example. The adversarial model has changed dramatically in 2026. LLMs and agentic tools have lowered the expertise required to find new 0-days, with close to no prior knowledge of the system being targeted. The whole software industry is feeling it. Mozilla reported around 20–30 security fixes per month for Firefox throughout 2025, rising to 423 in April 2026. Starting July 2026, Cisco introduced twice-monthly security disclosure windows, seven-day advance notifications, and planned quarterly hardening updates for major network operating systems. Google’s Threat Intelligence Group recently published an article in which they outlined how adversaries are leveraging multi-agent systems and LLM-driven pipelines to automate labor-intensive phases of vulnerability analysis, chaining capabilities together to reduce the window between patch disclosure and weaponized exploitation. Similar examples keep piling up and there is even a term for describing this turbulent era: the Vulnpocalypse.

The Tezos ecosystem is no exception. Over the course of 2026:

  • Six security upgrades were proposed and voted on by bakers on the Tezos L1 blockchain to address an ever-growing number of vulnerabilities. Most of these issues were threats to the liveness of the chain, but a few of them could have also put user funds at risk.

  • We shipped four releases of the Octez node to address flaws that put the health of the Tezos L1 network at risk.

As a consequence, we have had to invest significantly more engineering effort in addressing known (often still undisclosed) issues and proactively looking for new ones, to protect the chains and user assets. This shift in engineering focus is real, and it is here to stay. The Tezos bug bounty program, administered by the Tezos Foundation remains a valued source of community input on critical bugs. Yet models keep getting better. Safeguards implemented by frontier labs may not prevent every misuse. And new code keeps being written, bringing with it the possibility of new, increasingly subtle bugs with unexpected impact.

As security work grows, our engineers already spend more of their time working in private, preparing fixes that cannot safely be disclosed before they are deployed. The August 2026 Cosmos EVM incident illustrates why coordinating publication with deployment matters: according to Cosmos Labs’ post-mortem, a public pull request described the vulnerability and its exploitation path less than twelve hours before the first known attack. Six networks were ultimately exploited, even though patched releases were already available.

In addition, routine development can reveal sensitive information: a public branch fixing a seemingly harmless bug, or a discussion about the security of a future feature, can give an attacker a hint on a flaw in software already running on the network. Today, separating public development from private security work relies on every engineer recognising those implications before sharing a change or asking a question.

That is why we have decided to move day-to-day development to a private git repository starting December 2, 2026. Concretely, this means core engineers at TriliTech, Nomadic Labs and Functori working on Tezos L1 and Etherlink will push, propose and discuss their changes in a private Gitlab repository. We will set up automated mirroring of the master branch, meaning that our changes will be made available as soon as they are approved and merged. This change will give us more control over when sensitive changes become public. The software we develop will remain open source under the MIT license. We are committed to improving our documentation and tooling so that anyone can participate fully in Tezos ecosystem governance, including reviewing our changes and submitting protocol proposals. We also continue to welcome external contributions through the public monorepo. Any security vulnerabilities discovered by the broader technical community should also continue to be submitted via the Tezos Foundation’s established bug bounty process.

For almost a decade, the Tezos ecosystem has adapted to meet new realities. Security has been at the core of its design from the start, and systems built to last must be able to respond to threats their designers could not have anticipated. That now applies to how we develop it, too. The tools reshaping software security are not going away, and neither is our commitment to openness in what we ship. The Tezos L1 network and Etherlink are open source projects governed by the community. This change is about timing, not openness: it lets us fix problems before they are exposed, while ensuring these networks remain open source, auditable and governed by the community. Adaptability is a security quality, and it is one that has always characterised the Tezos protocol and its ecosystem. For our part, we will keep sharing what we learn as the industry adapts, and we welcome feedback from bakers, builders and contributors along the way.

3 Likes

Unfortunately, what you miss here is all of the nuance that comes with closing the doors of ideation and shifting governance to a rubber stamp. What position will bakers find themselves in? In this new paradigm you paint, bakers will be presented with a choice of act quickly to pass or leave the system open to whatever you just published. This equates to no choice at all. Governance is now shackled to rubber-stamping your developments. Developments that came from a black box and we have no choice but to trust is not what I signed up for. I would rather admit defeat here and now than try to patch away our standards and ethos.

I can easily boil down all of your arguments to; the Tezos experiment assumed being able to develop at human speeds, and in this new world it fails. Tezos doesn’t work anymore. This model is broken.

2 Likes

Hi. I’m Thomas Letan, an engineer at Nomadic Labs who has been working notably on Etherlink and Tezos X. I don’t want to misinterpret your message, but I am under the impression that we are potentially conflating two things here:

  1. The fact that changes by core engineering working at NL, TT and Functori will be reviewed in a private environment (which is something new), and

  2. Our practice to withhold the code for security sensitive code after they are being deployed (typically in the governance process of Etherlink).

The impact on the governance processes of the Tezos ecosystem blockchains cannot be answered uniformly in this case.

For “regular” upgrades, I want to challenge the idea that MRs on Gitlab have, up until now, been used as a coordination mechanism with bakers. Again, I don’t want to misrepresent your message, but in any case, we can answer to that: to the best of my knowledge, this isn’t the case, and for pretty good reasons. For bakers to be able to judge our proposals, they need proper documentation, they need clear changelogs, they need to be told up front what a protocol or kernel proposal will contain. If anything, we need to put more effort in strengthening these aspects of our ecosystem collaboration. It’s a recurring theme in recent discussions, I see that, and something we probably need to improve in the months to come. But we have to face up to the reality of this new environment.

For “security” upgrades, I can comment on something I’m very familiar with: the approach we are following now is the same we’ve been implementing for Etherlink since the very beginning. There were a few of them before this year, and there have been more in 2026, but in practice they are the same: they contain targeted patches, on top of the current version of Etherlink, we have uphold to the standard of communicating rigorously about each of them (sometimes with some delays, because of operational constraints), and even in the face of challenges brought by AI. In particular, these “best practices” haven’t changed and they won’t change. Etherlink does not have a security council, bakers remain in control and the process works based on the trust they have that our proposals only improve the security of the chain, and nothing more. The same applies to Tezos L1. Agents haven’t changed the nature of our collaboration, nor they have pushed us to request more unconstrained power. I think this is a good thing.

4 Likes

Thanks for the thoughtful breakdown, @lthms and welcome to the forum!

I’ll grant you the distinction between day-to-day private review and withholding security patches, as well as the operational reality of how bakers consume proposals today. On the mechanics of the current process, your points are fair. To be clear, my pushback isn’t a criticism of you or the hard work of any individual on the engineering teams.

Where we diverge is on what this represents in the bigger picture. Even if each individual step is a rational, necessary adaptation to a new threat landscape, it still forms part of a clear, cumulative pattern: a steady retreat from the open-by-default ethos that originally defined the protocol.

Whether the details of the workflow change how a specific patch is reviewed doesn’t really change the shadow this casts over the experiment as a whole. When the realities of the environment repeatedly force us to trade away foundational tenets just to keep the lights on, the writing on the wall is hard to ignore.

Asking these hard questions isn’t about pointing fingers. It’s because the answers need to inform every decision we make going forward. If we’re going to keep moving down this road, we have to find better ways to stay aligned.

1 Like

I’ve been saying for years that Tezos isn’t open source, it’s source available. This is just making it official finally. The doors it’s closing were never open in the first place.

1 Like

Thanks for the heads up. You guys keep doing the good work and I’ll press the vote button when the time comes to do my baker part. Over and out :saluting_face:

1 Like

Thank you @lthms (another individual engineer I have worked with over the years and admire) for stepping in to clarify the distinction between routine day-to-day review and withheld security upgrades. Even with that distinction drawn, the “AI / Vulnpocalypse” rationale in the announcement does not hold up to technical scrutiny. In my view, AI vulnerability discovery is a convenient pretext for the same steady withdrawal from public visibility and external accountability that has been underway across Tezos governance for the past two years.

1. Why the “Vulnpocalypse” Rationale Does Not Add Up Technically

The announcement states two things together:

“In addition, routine development can reveal sensitive information: a public branch fixing a seemingly harmless bug, or a discussion about the security of a future feature, can give an attacker a hint on a flaw in software already running on the network.”
“We will set up automated mirroring of the master branch, meaning that our changes will be made available as soon as they are approved and merged.”

Look at both halves of that workflow together:

  • Zero-day scanning: An AI agent (Claude Code, Codex, or a multi-agent pipeline) hunting for 0-days on live Tezos L1 or Etherlink nodes scans the deployed code and master, both of which remain completely public.
  • Known security fixes: When engineers do recognize that a bug or fix is security-sensitive, standard open-source security hygiene (used by the Linux kernel, Chromium, Rust, Go, Ethereum clients, and Cosmos after the August 2026 post-mortem cited above) is to develop that patch in a private security advisory fork and withhold publication until coordinated deployment (which, as @lthms noted above and as today’s Ganesha 7.3 post-mortem in #7243 illustrates, is already how targeted security patches are handled).
  • “Seemingly harmless” routine bugfixes: Conversely, if an engineer and reviewer do not realize a routine fix touches a live vulnerability, under this new workflow they approve and merge it, and automated mirroring immediately publishes that exact commit diff onto public master weeks or months before node releases ship or L1 protocol upgrades activate. Automated patch-diffing of newly merged master commits is the exact AI attack vector described in the Google Threat Intelligence report cited in the OP.

You do not need to move 100% of routine day-to-day L1 protocol, node, RPC, and Etherlink engineering behind closed doors to keep security advisories private, and auto-mirroring master means unrecognized bugfixes still land in public before deployment anyway.

Look at how open-source security handling actually works in practice today (20 November 2026): this morning we shipped two security releases in Taquito v25.1.2 and v25.1.3 (also rolled into v25.2.0-rc0): GHSA-9wjf-4grc-2g2h in @taquito/local-forging (reported directly to security@ecadlabs.com by @byfor8) and GHSA-6r9w-cxjj-5776 in @taquito/sapling (reported by Rahul Saxena of Bluethroat Labs and forwarded to us by security@tezos.com today). Once we received the @taquito/sapling report this morning, we opened tested fixes within an hour and published v25.1.3 to npm in ~2.5 hours (verified by the reporter against his Octez 25.3 replay harness), all while keeping day-to-day Taquito and Beacon engineering 100% open on GitHub.

2. What Moving tezos/tezos Private Actually Locks Out

In Post #3, @lthms wrote: “For ‘regular’ upgrades, I want to challenge the idea that MRs on Gitlab have, up until now, been used as a coordination mechanism with bakers.”

Bakers pressing a vote button on a finished proposal (as @Primate411 noted in Post #6) are not the only external participants in an open-source blockchain:

  1. Independent SDKs, wallets, remote signers, and indexers lose advance visibility: Having maintained Taquito (and previously Signatory) across all 21 on-chain Tezos protocol upgrades since Athens in 2019, watching in-flight MRs and technical discussions on gitlab.com/tezos/tezos is specifically how external infrastructure teams catch breaking RPC schemas, binary-encoding shifts, and gas/storage parameter changes before they land as a fait accompli on master. That early visibility became our last remaining coordination surface after our team was dropped from proto-ecosystem calls and disconnected from the shared Slack-Connect channels on September 4.
  2. It forecloses the original 2014 promise of open protocol governance: The announcement states: “We are committed to improving our documentation and tooling so that anyone can participate fully in Tezos ecosystem governance, including reviewing our changes and submitting protocol proposals.” In the 2014 Tezos Position Paper and White Paper, self-amendment was designed so Tezos would never be captured by a single closed core development team, allowing independent teams to compete on protocol proposals and earn on-chain invoices on merit. Across more than 8 years on mainnet and all 21 on-chain governance upgrades (through 025 Ushuaia), zero proposals from outside the Foundation-funded core-contractor pipeline have ever passed, and zero protocol invoices have been paid since Lima in December 2022. How is any external team supposed to collaborate on or propose protocol changes when the active development tree, code review, and design discussions live on a private GitLab server accessible only to Trilitech, Nomadic Labs, and Functori?

3. Not About AI: A Consistent Retreat from Public Accountability

If this change were really about AI agents, it would not fit so neatly into a pattern that began long before any “Vulnpocalypse” report. Every other surface through which token holders, bakers, and independent builders could evaluate core execution and spending was already closed using other justifications:

  1. Public issue tracking and roadmaps went private (Asana → Linear): Backlog triage, roadmaps, and sprint prioritization moved off public GitLab/GitHub trackers into Asana and then into a private linear.app/tezos workspace.
  2. Core-entity financial transparency was removed from the Activity Reports (July 2024): The Tezos Foundation removed the table disclosing funding allocations to core development entities (Nomadic Labs, Trilitech), highlighting $10.30M in newly approved external grant initiatives at the front of the two 2025 reports while disclosing $89.06M in actual payouts to grantees and service providers on the back page without itemizing where the other $78.76M goes.
  3. Independent infrastructure teams were consolidated out of the ecosystem: Over the past few years, independent maintainer organizations (Cryptonomic, Blockwatch, AirGap, and now ECAD Labs, among others) saw their Tezos roles wound down, absorbed in-house, or defunded; critical baker infrastructure like Signatory had its maintenance funding cut (#7093); and actively maintained open-source SDKs (Taquito and Beacon) were forked under trilitech/* as octez.js and octez.connect.
  4. Cross-team coordination channels closed: Independent maintainers were removed from proto-ecosystem calls and shared Slack channels.
  5. And now (December 2, 2026), day-to-day development on gitlab.com/tezos/tezos goes private: External builders, wallets, indexers, bakers, and token holders are left with post-merge code drops on master.

None of steps 1 through 4 had anything to do with AI vulnerability scanners. What they all share is the governance preference @murbard stated candidly on this forum in #7093/13 back in June:

“Do teams at TT or NL sometimes not do a good job, waste resources, not deliver. Of course. But when they do, teams can be reorganized and we don’t have to deal with online disparagement campaigns…”

When day-to-day branches, MRs, review timelines, and design debates on gitlab.com/tezos/tezos are moved onto a private corporate server, nobody outside the house group can see stalled work, false starts, wasted resources, or breaking changes in real time; the public only sees what is curated onto master after the fact.

With day-to-day git development, issue backlogs, coordination calls, and core-entity budget breakdowns all moved behind closed doors, forum.tezosagora.org is now the last surviving public accountability surface in Tezos governance. Which legal entity currently owns and operates tezosagora.org, and will leadership commit on the public record to keeping this forum open, indexed, and unpruned?

2 Likes

Hi Jev, thanks for asking, we at TCF have owned and operated Agora, and have always kept Agora open and all the things you have mentioned. As long as people follow the rules that we set at the inception of Agora, then all is well. If you have any suggestions, we’re all ears.

2 Likes