After hearing about the nearly 4,000 BTC (approximately $320 million) Liquid exploit on September 6, I pointed MASH, Alpen Labs’ internal Multi-Agent Security Harness, at the Elements codebase. We’ve been working on an agentic red team to continuously secure our own ecosystem, so I was curious to see what these tools could do.
In about an hour, the agents had traced the root cause and written a post-incident report that included a local reproduction of the exploit.
But their finding didn’t match the speculation circulating online. The prevailing explanation was that a public commit had exposed an existing vulnerability before nodes could upgrade. Our agents pointed to a different cause: a change intended to fix an older bug had introduced a new flaw in the cache key.
That surprised me. After closely reviewing the findings and the local reproduction, I shared our reports with Adam Back and the Blockstream team.
Blockstream confirmed that the root cause matched its internal findings. That night, I used MASH to support its head of engineering with further investigation, fix development, and validation. The subsequent remediation report documents that work.
From code review and audits to continuous monitoring and penetration testing
Much of the conversation around cyber-capable AI focuses on how fast attackers can find and exploit weaknesses. We don’t have evidence that AI was used in the Liquid attack, but we saw firsthand what it could do for defense.
A blockchain is a state machine. Transactions change its state; upgrades can change the rules, and new applications and integrations change how those rules are exercised. Security assumptions need to keep holding as the system evolves.
Audits and code review provide essential checkpoints. But securing live blockchain systems as AI capabilities advance requires ongoing testing beyond those checkpoints. Teams of agents coordinating through harnesses like MASH can monitor the live ecosystem, reproduce suspicious behavior in controlled environments, and surface findings before a bug turns into an exploit.
How we’re applying MASH at Alpen
With Alpen approaching mainnet, we’re working toward a 24/7 MASH red team for our evolving ecosystem.
In staging, agents test the bridge, proof system, consensus, and access controls in parallel. They share findings and triage results, using independent models for additional validation and local reproductions. Eventually, we’d like to extend coverage across the ecosystem.
Agents continuously surface potential attack paths and supporting evidence. Alpen contributors remain accountable for fixes and decisions about what ships.
We built MASH through our partnerships with OpenAI and V12. It combines publicly available frontier and open-weight models, restricted cyber models through OpenAI’s Trusted Access for Cyber program, and specialized models such as Zellic’s v12.
Designing fault-tolerant systems
We can’t expect MASH to find every bug before an attacker does. On Alpen, finding a bug should still leave an attacker facing challenge periods, withdrawal-rate checks, and further validation and authorization before deposited BTC can be withdrawn. If one defense fails, other independent defensive layers still have to hold.
Those protections also need to let valid withdrawals complete and the chain continue finalizing state. An attacker shouldn’t be able to turn a safety check into an indefinite freeze. Safety and liveness both matter when we can’t rely on rolling back a finalized Bitcoin transfer. We’ve built Alpen’s bridge peg mechanism around these principles.
I want MASH testing that fault tolerance as Alpen evolves. I think the same work should become routine across Bitcoin and other ecosystems.
The incident report was prepared on September 6, within hours of the exploit. I’m sharing it now, alongside the subsequent remediation report, after delaying public release for security reasons.