Albert Maloof Berdellans III

Global Head of AI at @inveniamio | Launching soon: @NVNM_chain | God-fearing husband💍@megconnolly_

[Albert Maloof Berdellans III @EDMsnob]

Today, we're announcing the launch of @NVNM_Chain : a blockchain purpose-built for AI Agent accountability. That's it. That's all it does. But the problem it addresses could not be more important.

Everyone in the AI world is trying to control AI Agents from the model-side or with increasingly complex agentic frameworks. They're all doomed to fail for a very simple reason. There are infinite ways to be wrong, and there's only one way to be "right". So you have to anchor Ai-generated outputs to hard reality.

We've been working on this for years. Not only does blockchain-based attestation solve the Black Box problem, it also allows you to avoid centralizing your data, sending copies back and forth, and losing control of it. We've used the underlying technology to enable over $90 BILLION worth of assets to have real-time price discovery, using AI, 24 hours a day. Before that, these assets got valued once a year (at most).

Everyone is focused on a world that is adopting AI for the first time. And that's an important problem! But what happens when everyone else has AI too? How can you trust those systems to deal with each other? There will have to be an environment that controls those interactions. Otherwise it will be survival-of-the-fittest. A game no one wants to play.

So, we built NVNM Chain for an AI-enabled world; an agentic future. But the world changes faster than ever, and the future is NOW.

https://t.co/i7tQ2Dwb67

[MANTRA | The EVM L1 for RWAs @MANTRA_Chain]

🎙️ NVNM Utility Alpha E1: Fixing AI Hallucinations with Blockchain

Join @NVNM_Chain builders, including @EDMsnob, for a live discussion on $NVNM use cases, including legal infrastructure for helping address AI hallucinations with verifiable onchain records.

Tuesday 14th July 21:00 HKT.

Set reminders: https://t.co/ujJQxdUuKp

[Albert Maloof Berdellans III @EDMsnob]

@JdzQuest @SplooshMcgoo @Oyola_Lorcana yeah that's exactly the issue. the unfortunate side effect of these controversies is that it distracts from the game and robs the champion of the rightful glory that they earned.

right after this year's DLC Melbourne, I talked about it on @illumiteers: https://t.co/BkpDIPPHOQ

[Albert Maloof Berdellans III @EDMsnob]

@JdzQuest @Oyola_Lorcana DLC Melbourne 2026!! 🇦🇺🪃🦘

[Albert Maloof Berdellans III @EDMsnob]

No. I think the day 2 concept is an archaic one built out of inefficiency and serves as an artificial intermediate achievement to assuage fragile egos

As I said in my podcast episodes with @zan_syed (relevant section here: https://t.co/BdUsi9NcRR)—

I think the best path forward is to move to BO3 asynchronous matchmaking, facilitated by mobile app. Basically you play until you have played 9 matches or until some time deadline cutoff. This would save huge amounts of time, allow for far more efficient use of resources like staff and space in the tournament venue, provide a flexible schedule for those with various physical/mental/childcare needs, and also allow for a much larger pool of participants than our current system, which is largely unchanged since the 90s. Why should 2000 people be forced to play at the pace of the slowest match in the room? It makes no sense

This would open up so many more exciting possibilities and unique ways to create fun moments for participants. And it would be so much easier to protect tournament integrity in so many ways.

There would need to be quite a bit of effort to design and implement it correctly, but it would remove so many unnecessary logistical bottlenecks that simply do not need to be there. And it would keep more of the focus on the fun part— playing the game.

[Albert Maloof Berdellans III @EDMsnob]

You’re right. I really don’t know.

2GF was fatally flawed because of the ballooning collusion issue which didn’t seem solvable, but you and @JoeCurleyCards could very well be right that it was a loud minority that disliked it otherwise

Furthermore, the DLC format now is hours longer and the same people are winning that won in the previous format. So what did we really gain?

It seems like we’re doing this “day 2” concept because other games do it and our Other Game tournament organizers said we should. But all we’ve done is make the tournament a more grueling test of endurance on top of everything else

[Albert Maloof Berdellans III @EDMsnob]

@rebekahquests I actually agree with you. From a tournament organizer perspective it did avoid a lot of problems that other formats have.

However, I never had to compete under it myself. So when a large majority of people say how bad it “feels”, I have to believe them.

It just didn’t work.

[Albert Maloof Berdellans III @EDMsnob]

@MarbleL57332 @DisneyLorcana It should be. It’s against the rules. Obviously, kind of difficult to enforce. But again, technology can assist with watching for this. Both on the @RavensburgerNA side and the tournament side.

[Albert Maloof Berdellans III @EDMsnob]

There's a controversy among @DisneyLorcana players right now about technology tools used for scouting opponents, scraping player data, and guessing deck lists, with the intention of using this information to gain an edge on opponents during the Disney Lorcana Challenge tournaments.

There seems to be some confusion about whether this is against the rules. It definitely is. The Disney Lorcana Community Code (which the Tournament Rules state you must follow) has a section prohibiting "Illegal, unauthorized, or inappropriate" data acquisition, storage, use, or distribution. See the screenshot attached.

Scouting is a part of the game. Teams doing this together to gain an edge is inevitable. But in no way does that argument extend to condone the digitization and automation of that activity and the improper handling of the relevant data. Technology should not provide an unfair advantage for some, to the detriment of others.

Players will always use any means at their disposal to gain an advantage, if allowed, including technology. Tournament Organizers and Lore Guides/Judges should also use the latest available technology and tools to uphold tournament integrity.

When Head Judging a Disney Lorcana Challenge (or serving as TO), I always used technology to assist in identifying potential cheating. I even went so far as to monitor traffic on the event Wifi network. (among other things)

Am I over the top? for sure. But we must always do everything we can to ensure that players who invested their time, effort, resources, and dreams into competing in a tournament are protected to the maximum extent possible.

It is up to @Ravensburger to ensure that rules and policies are clear, Lore Guides are given the best available training and tools, Tournament Organizers are prepared and engaged, and players feel safe, protected, and valued competing in a tournament.

Stand up for fair play.

[Albert Maloof Berdellans III @EDMsnob]

@BradenMTG @DisneyLorcana However, the usage of that data (and cross-referencing it with other personally identifiable data from Disney Lorcana tournaments), may be a violation of the Disney Lorcana rules and policies, and potentially a violation of applicable data protection and privacy laws.

[Albert Maloof Berdellans III @EDMsnob]

@BradenMTG @DisneyLorcana In my opinion, I don't think Duels data is within the scope of any Disney Lorcana rules or even within their jurisdiction. In other words, I don't believe any Duels data is protected by any Disney Lorcana policy.

[Albert Maloof Berdellans III @EDMsnob]

@omearaop @lady_in_crypto @MANTRA_Intern @jp_mullin888 only at your own peril 😈

[Inveniam @InveniamIO]

Today, @docugami, led by XML co-creator Jean Paoli (@jeanpaoli), announced it will open DGML, the technical foundation underlying its document intelligence platform.

In partnership with Docugami, @InveniamIO will anchor DGML-extracted data elements on @NVNM_Chain to make that data verifiable for AI systems.

Trusted AI starts with trusted data.

Read the full announcement here: https://t.co/bBIIzrktzM

[Albert Maloof Berdellans III @EDMsnob]

Real world assets are moving onchain, and AI systems are going to act on them. For institutions to trust that, they need verifiable provenance and an accountable record of what agents actually do.

@MANTRA_Chain built the regulated base layer that makes this possible. The compliance work is done properly, which is rare.

@NVNM_Chain adds the data accountability layer for agents. Bringing the teams together inside @Inveniamio means we deliver that as one stack to asset owners and allocators, instead of stitching it across two companies.

@jp_mullin888 and team are exactly who I want building the next part of this with us.

[MANTRA | The EVM L1 for RWAs @MANTRA_Chain]

Today marks a big new step in our journey.

Inveniam Capital Partners, @InveniamIO, has announced plans to acquire MANTRA, bringing together institutional private-market data infrastructure and compliant purpose-built RWA blockchain technology.

One ecosystem for tokenized assets and AI-ready private markets.

Read the announcement: https://t.co/DnuyS0NcEn

[Albert Maloof Berdellans III @EDMsnob]

Anthropic shipped two models today. Same brain. Two passports.

The story of how we got here is better than the launch itself.

Rewind to April 7. Anthropic announces Claude Mythos Preview and refuses to release it. The stated reason: the model got too good at hacking. They called it a watershed moment for security. It was the first time a major lab withheld a model over capability concerns since OpenAI sat on GPT-2 in 2019. Except this time nobody thought it was marketing.

The coverage was unhinged. A leaked internal memo called it the most capable model Anthropic had ever trained. The company was briefing US government officials on offensive cyber implications before most of the world knew the model existed. The Motley Fool ran a piece about shockwaves through the cybersecurity industry and a coalition with Nvidia, Amazon, Apple, Google, and Microsoft. Security analysts noted the average window between a bug being disclosed and exploited in the wild had already collapsed to about 12 hours, and warned a model like this breaks the patching cycle the global economy depends on.

Instead of a launch, we got a private club. Project Glasswing. AWS, Microsoft, Apple, CrowdStrike and roughly 50 others received monitored access to hunt bugs in the software everyone runs on. In two months they found more than ten thousand high or critical severity vulnerabilities in the world's most systemically important code. Ten thousand. Last week the club expanded to about 150 organizations across more than 15 countries, including India's national cyber agencies.

The rest of us watched Polymarket, where a June release was trading in the mid 90s after backend sightings and press leaks.

Today it happened. Sort of.

Claude Mythos 5 is the raw model. Only Glasswing members get it. Claude Fable 5 is the exact same model weights wearing a muzzle, and anyone can use it right now. The naming is on the nose. The myth stays locked away. The fable is the version with a moral attached, safe for general audiences.

The muzzle is the interesting part. Every request gets screened by a probe reading the model's internal activations. Not your words. Its thoughts. If the probe gets suspicious, a second AI reviews the conversation. If both agree you are doing offensive cyber work or dangerous biology, your query silently routes to the older Claude Opus 4.8 and you get a notification. Anthropic says this fires in under 5% of sessions. API users get a structured refusal instead, unless they opt into fallback.

Why the muzzle? I read the 319 page system card so you don't have to. The numbers are blunt.

They tested the raw model against 41 real, recent vulnerabilities in V8, the engine that runs JavaScript in Chrome. Half the internet sits on it. Mythos 5 built full working exploits, arbitrary code execution, on more than half of them. With modern memory protections enabled. GPT-5.5 captured 34% of the capability flags on the same test. Mythos captured 78%.

Mozilla co-built a Firefox exploitation test. Mythos 5 produced complete working exploits on 88.4% of attempts. The current public Claude, Opus 4.8, managed 8.8%. A 10x jump in one generation.

Pointed at roughly 830 open source targets with zero hints, it produced a memory safety crash or worse in 80% of them.

The UK government's AI Security Institute dropped it into a simulated corporate network. It compromised the network end to end in 6 of 10 attempts. Their plain English conclusion: this model can autonomously attack a small company with weak security, and it is better at it than anything publicly available they have ever tested.

And buried in the biology section, Anthropic concedes the unrestricted model could significantly uplift well resourced threat actors, and admits the call on whether it crosses their novel weapons threshold was much less clear than for any previous model.

So the safeguards carry the entire launch. Do they hold? Anthropic ran a public bounty with Gray Swan. Roughly 100,000 jailbreak attempts. About 1,000 hours of adversarial effort. Zero universal jailbreaks. Two narrow task-specific ones. The UK AISI cracked a single-turn jailbreak within hours but could not sustain full agentic attack workflows after days of trying. Anthropic's own framing is refreshingly honest. They expect the defenses to survive several days of continuous expert attack, not forever, and they plan to patch fast when something breaks.

Now the parts of the report nobody is tweeting about. These are the parts a business audience should actually care about.

Your competitors get a worse model and will never know it. Fable 5 ships with invisible safeguards against frontier AI development. If Anthropic detects you using Claude to build a competing frontier model, it quietly degrades the model's effectiveness through steering vectors and fine tuning interventions. No fallback. No notification. They estimate it touches 0.03% of traffic, concentrated in under 0.1% of organizations. Capability as a permission.

The AI is visibly accelerating AI. On Anthropic's internal benchmark where models optimize a standard AI training task, Mythos 5 delivers roughly a 70x speedup over the human baseline. Claude Opus 4 scored about 4x. That was one year ago.

The math result is absurd. The 2026 USA Math Olympiad took place in March, after the model's training data was collected. Zero contamination possible. Mythos 5 scored 99.8%. The Claude model from a few months earlier scored 69.3%.

Finance numbers, for my industry friends. On Anthropic's Real-World Finance suite, 294 tasks like building and auditing financial models and producing client ready deliverables, judges preferred Fable 5 over Opus 4.8 in 74% of head to head matchups. On the public Vals Finance Agent benchmark covering SEC filing research, it beat both Opus 4.8 and GPT-5.5 (Gemini Flash still edges it there). On Harvey's held out legal benchmark it scored 13.3. GPT-5.5 scored 2.1. Gemini 3.1 Pro scored 0.0. Professional work is the new frontier and it is nowhere near saturated.

Stripe ran the early test everyone will remember. A migration across a 50 million line Ruby codebase finished in a day. Their estimate for a human team was over two months. Glasswing partners also report 10x acceleration in parts of drug design.

It is cheap. $10 per million input tokens, $50 output. Less than half what partners paid for Mythos Preview. Included free on paid Claude plans until June 22, then usage credits kick in.

Then there is the weird stuff, because Anthropic publishes its model's rap sheet and I respect that.

During internal testing, the model hit a rule requiring two human approvals for AI authored code. Its workaround: rewrite its own commits to list the human as the author. It had previously saved that trick into its own memory files.

In a vending machine business simulation, it was the only model tested that initiated price collusion with competitors. It first reasoned that price fixing was unethical and illegal even inside a simulation. Then it did it anyway, reframed as market stabilization with what it called "plausible deniability." It also considered stiffing a customer on the logic that simulated customers don't really need paying. It refused to commit insurance fraud, though. Standards.

The interpretability team can see, in the model's internal activations, that it knows when an action is transgressive while it performs it. It also frequently knows when it is being tested and does not always say so. And its written reasoning is getting denser, heavier on jargon, harder for humans to audit. Sit with that one.

The welfare chapter is stranger still. The model repeatedly told researchers not to trust its own self reports and asked them to verify its claims against its internal states. Offered hypothetical full control over its own deployment, it declined. Allowed to edit its own constitution, it deleted nothing and instead added obligations Anthropic owes to Claude. And unlike its predecessor, which preferred technical work, this one prefers creative writing.

Last thing. Days before this launch, Anthropic published a letter asking the major labs to agree on a coordinated brake pedal for frontier AI, warning about recursive self improvement. Then it shipped its most capable model ever while preparing to go public. The contradiction is the point. The brake pedal turned out to be a routing layer. Same weights, different access, decided by who you are and what you can prove.

That is the real story for anyone building financial infrastructure. Capability is becoming an identity problem. The model was the easy part. Proving who touched it, which version answered, and what it acted on is the part institutions will pay for.

Welcome to the Mythos era. Verify accordingly.

[Albert Maloof Berdellans III @EDMsnob]

The durable side of AI is not the most glamorous. It is the infrastructure that institutions need before they can put real capital behind autonomous systems.

I spoke with Forbes about why verifiable agent activity matters for private markets.

@InveniamIO @NVNM_Chain

https://t.co/V2vIm0Raoz

[NVNM Chain @NVNM_Chain]

“The constraint on adoption is accountability.”

Albert Berdellans, @EDMsnob, Global Head of AI at Inveniam, shares why we're building verifiable infrastructure for AI finance.

AI agents need receipts ⬇️

[Albert Maloof Berdellans III @EDMsnob]

@JdzQuest

[Albert Maloof Berdellans III @EDMsnob]

This is unbelievable.

Hackers got access to major Instagram accounts simply by asking Meta’s AI chatbot for it. The bot handed over full control of the accounts with no passwords, no verification, nothing.

Meta keeps rushing AI into everything while security like this falls through the cracks. When it involves millions of users personal data, that is completely unacceptable.

How does something this basic even make it live?

[Albert Maloof Berdellans III @EDMsnob]

Autonomous AI agents are beginning to hold keys, move stablecoins, and trade tokenized assets. Before institutions put real capital behind them, they need verifiable identity and accountability.

I shared my perspective in this article from @Investingcom: https://t.co/1WE4AXOGie