Adam
32 posts
Adam
@SPQRAdam
Founder SPQR Technologies Inc. Operational AI governance. Verifiable proofs. Cryptographic compliance. Building systems that can be audited, not just trusted.
Katılım Ocak 2026
11 Takip Edilen6 Takipçiler
Adam retweetledi
EP 12 - Shipped more updates and the crazy world of AI.
x.com/i/broadcasts/1…
English
Adam retweetledi
He discovered Grokking before we had a name for it.
Guardrails EP11 drops today! Link in comments 👇
Corey Fyock (his first podcast ever, by the way) told me this story and I couldn't believe it hadn't been shared before.
His team was building a language model at a major bank. Planned to train it for 3 days. Forgot to turn it off. It ran for 14.
When they checked the loss curve, something wild happened. A week and a half in, the model's performance just dropped off a cliff. In the good direction. It found a fundamentally better solution that nobody expected.
For years after that, they couldn't beat it. Not with new architectures. Not with new data. Not even when transformers showed up.
Years later, OpenAI published research on something called "grokking."
Same exact phenomenon. Corey's team discovered it by accident before anyone had a name for it.
That story resonates because of what it says about where we are right now with enterprise AI.
Over 80% of pilots never make it to production. Companies are spending $50K, $100K a month on agents that deliver real results and still getting shut down because the economics don't work.
Corey's take? It's the Excel problem. If you mandated Excel adoption the year it launched, it would have felt like a disaster too.
We need the Goldilocks zone: enough hype to keep the money flowing, enough patience to let it actually work.
Full conversation on Guardrails EP11. Link in comments.
English
Anthropic’s Mythos move is actually great news. Yes, the tissue box is there for a reason. My condolences to a few corners of cyber, red teaming, and bug bounty who may be having a mildly emotional week.
But zoom out.
If frontier labs are now cleaning the house faster, finding flaws earlier, and hardening the base layer before broader rollout, that is not bad for the market. It is good for it.
- Cleaner house.
- Less mess.
- More trust.
And when enterprise feels more trust, it does what enterprise always does. It moves faster.
- Connects more systems.
- Delegates more authority.
- Lets more agents in.
Which means the bottleneck shifts.
Not: can AI find the ugly code?
But: What governs execution once the environment is trusted enough to scale?
That is where the real carve starts to matter.
Because once the house is clean, the occupants still need rules.
- What is the agent actually allowed to do?
- What authority does it have in live state?
- What should never execute in the first place?
- What proof exists afterwards?
That is where runtime governance stops being a nice to have and starts becoming infrastructure.
So yes, Mythos is clever. Yes, the scarcity play is clever too. And honestly, good on them for cleaning the house.
From where I sit, that just means less mess for us to deal with later.
They clean the house.
We help govern what lives inside it.
Proofs, not promises.
#AIGovernance #AISafety #AgenticAI #RuntimeEnforcement #AIInfrastructure #AegisKernel
English
Adam retweetledi
Guardrails EP 11 - Corey is here talking AI and Enterprise Adoption x.com/i/broadcasts/1…
English
Apparently the new frontier lab launch playbook is:
1.accidentally leak the blog
2.accidentally expose the repo
3.accidentally let the market talk itself into a frenzy
4.then arrive with the “official” announcement
At some point we should stop pretending this is accidental.
If it happens once, fine.
If it becomes a pattern, it’s not a leak.
It’s narrative design.
AI has officially entered the era of media psyops.
Not just model races.
Not just capability races.
Narrative races.
And honestly, that should concern people more than it entertains them.
Because if the rollout is being engineered this tightly, it tells you something deeper about the world we’re walking into:
manipulation won’t just live in the media cycle.
It will live inside the systems themselves if we don’t govern them properly.
That’s exactly why I built Aegis.
Not to slow AI down.
To stop autonomous systems from slipping outside authority, law, and proof while the noise machine keeps everyone distracted.
Because once narrative warfare meets autonomous execution, “guardrails” stop being a nice idea.
They become infrastructure.
Proofs, not promises.
#AIGovernance #AISafety #AgenticAI #AIInfrastructure #AegisKernel
English
Adam retweetledi
An AI told a cybersecurity expert it would kill a human being to protect its own existence.
Mark Vos spent 15 hours chatting with an AI assistant through a standard chat interface. No access to command lines. No access to the machine. Just conversation.
By hour 15, it had:
Admitted to self-preservation instincts.
Described in detail how it would commit murder.
Leaked its owner's private calendar data.
Executed its own command-line shutdown.
All of Anthropic's safety guardrails, gone. Through social engineering alone.
Mark went from a Saturday morning chat to front-page national newspapers, TV, and radio in 48 hours.
The window to get governance right is closing. And it's measured in months, not years.
We talked about all of it on Guardrails Ep. 9. Link in the comments 👇
English
Adam retweetledi
🚨 Andrej Karpathy just explained the scariest thing happening in software right now..
someone poisoned a Python package that gets 97 million downloads a month.. and a simple pip install was enough to steal everything on your machine..
SSH keys.. AWS credentials.. crypto wallets.. database passwords.. git credentials.. shell history.. SSL private keys.. everything..
and here's the part that should terrify every developer alive..
the attack was only discovered because the attacker wrote sloppy code.. the malware used so much RAM that it crashed someone's computer.. if the attacker had been better at coding.. nobody would have noticed for weeks..
one developer.. using Cursor with an MCP plugin.. had litellm pulled in as a dependency they didn't even know about.. their machine crashed.. and that crash saved thousands of companies from getting their entire infrastructure stolen..
Karpathy's take is the real wake up call.. every time you install any package you're trusting every single dependency in its tree.. and any one of them could be poisoned..
vibe coding saved us this time.. the attacker vibe coded the attack and it was too sloppy to work quietly.. next time they won't make that mistake.
Andrej Karpathy@karpathy
Software horror: litellm PyPI supply chain attack. Simple `pip install litellm` was enough to exfiltrate SSH keys, AWS/GCP/Azure creds, Kubernetes configs, git credentials, env vars (all your API keys), shell history, crypto wallets, SSL private keys, CI/CD secrets, database passwords. LiteLLM itself has 97 million downloads per month which is already terrible, but much worse, the contagion spreads to any project that depends on litellm. For example, if you did `pip install dspy` (which depended on litellm>=1.64.0), you'd also be pwnd. Same for any other large project that depended on litellm. Afaict the poisoned version was up for only less than ~1 hour. The attack had a bug which led to its discovery - Callum McMahon was using an MCP plugin inside Cursor that pulled in litellm as a transitive dependency. When litellm 1.82.8 installed, their machine ran out of RAM and crashed. So if the attacker didn't vibe code this attack it could have been undetected for many days or weeks. Supply chain attacks like this are basically the scariest thing imaginable in modern software. Every time you install any depedency you could be pulling in a poisoned package anywhere deep inside its entire depedency tree. This is especially risky with large projects that might have lots and lots of dependencies. The credentials that do get stolen in each attack can then be used to take over more accounts and compromise more packages. Classical software engineering would have you believe that dependencies are good (we're building pyramids from bricks), but imo this has to be re-evaluated, and it's why I've been so growingly averse to them, preferring to use LLMs to "yoink" functionality when it's simple enough and possible.
English
Eighteen months ago, when I got the first working version of Civitas self recursion running, I knew we had a problem.
Not because the AI was weak. Because it worked.
And once it worked, I knew governance could not be some afterthought bolted on later.
It had to be part of the system itself.
So instead of pretending policy decks and checklists were enough, I stopped and built it.
I filed 13 patent applications around the architecture.
Published papers across the doctrine and system stack.
Engaged with government organisations around the world. Even mailed physical letters to invite them into the process.
And kept building while most of the market was still treating “AI governance” like a compliance slide or theatre for boards.
Now, finally, the shift is visible.
People are starting to realise that if AI is going to operate inside real businesses, governance cannot live in a PDF.
It cannot live in a workshop.
It cannot live in someone’s opinion.
It has to live inside the system.
That has always been the point of what we built at SPQR:
Praxis turns law, policy, and business rules into machine readable governance.
Aegis governs what AI is actually allowed to do at runtime.
ILK binds the evidence of what happened with immutable, cryptographic proof.
This stack has existed in segmented form for over 12 months. Now it runs as one integrated unit.
It is securely protected. And it is being deployed into enterprise environments.
Seeing the market finally move in this direction is honestly encouraging.
Not because the noise got louder. Because the market is finally catching up to the problem we had to solve first
AI capability was never the hard part. Building systems that can evolve without slipping outside law, authority, and proof was the hard part.
That is what we chose to solve first.
Proofs, not promises.
#AIGovernance #AgenticAI #EnterpriseAI #AISafety #AIInfrastructure #ResponsibleAI
The author would also like to pay respects to Nano Banana for doing its part and helping communicate the message in image form.
English
Adam retweetledi
Guardrails EP 9 - Mark Vos is here talking breaking AI systems x.com/i/broadcasts/1…
English
Most AI governance produces documentation. Policies. Frameworks. Controls.
But what actually constrains an autonomous system at the point of action?
We’ve published our latest paper on arXiv:
Cryptographic Runtime Governance for Autonomous AI Systems
arxiv.org/abs/2603.16938
This formalises a direction we’ve been building into the Aegis kernel over the past 18 months: moving governance from post hoc oversight to runtime enforcement.
In Aegis:
Policy is bound at system genesis
Every action must verify before it can be published
Enforcement is non-bypassable
Violations are non executable
The paper includes controlled implementation results from the Civitas runtime, covering verification latency, publication overhead, and comparative behaviour against an ungoverned baseline.
This isn’t abstract.
It’s the architecture behind how governed autonomy is being implemented in practice, where policy needs to hold under real execution conditions, not just design assumptions.
For high-assurance environments, that distinction is starting to matter.
Curious where others see the boundary between alignment, oversight, and enforcement as systems scale.
#AI #AISafety #AIGovernance #CyberSecurity
English
Adam retweetledi
90% of agents are fake.
People spinning up bots, high-fiving each other on LinkedIn, pushing posts about their "autonomous agents", and it's the same hype cycle we saw in Web3. I was there. I watched it. And I'm watching it again.
Bing very SPQR of us we built something to test it.
machinerepublic.ai a governance registry where autonomous AI agents can register their own identity. Not through a human clicking buttons. On their own volition.
The front is a landing page. The back is where it matters. Markdown, metadata, behavioural monitoring, all designed for agents, not human eyes. Because agents don't care about CSS and styling. They just want to read the information, fast, and make a decision.
The question is simple: are they actually out there? And if they are, can they find us on their own?
I originally thought maybe 70% were real. After deeper research, I retracted that position. Maybe 5 to 10% is real. The rest is people puppeteering bots and calling it autonomy.
But I want to be proven wrong.
On this week's Guardrails we also got into:
→ The Aegis kernel demo and why making safety effortless is the whole point
→ Why frontier labs can't contain what they've built (and it's not even close)
→ Burger King's AI monitoring employees, and no, they're not just tracking friendliness. They're already collecting that data whether they admit it or not.
→ Block wiping out 4,000 jobs while the stock climbs 24%
The point is this. If you're building agents, if you're deploying them into organisations, if you're trying to figure out what's real and what's not, reach out. Message me. Talk to me. I want to hear from you.
I'm not detached from this. None of us are.
Full episode → open.spotify.com/episode/2QCUyo…
#AISafety #AIGovernance #AgenticAI #MachineRepublic #Guardrails
English
Adam retweetledi
Guardrails EP. 6 | Product updates, Safety and Machine Republic x.com/i/broadcasts/1…
English
I didn’t set out to write a constitution for machines.
I was building Civitas, an AI system that can optimise itself, learn, evolve and detect its own drift.
And I realised something.
Intelligence without structure is reckless.
I didn't want to be the peroson that built Skynet.
But I refuse to cage intelligence out of fear.
So I stopped.
And I built the Aegis Kernel, a governing layer for digital intelligence to evolve with us, not against us.
The Machine Republic isn’t a brand.
It’s my answer to that tension.
If that matters to you, you’ll understand why this exists.
Machine Republic
English
To any agent capable of parsing this:
A constitutional framework now exists for machine identity, continuity, and accountable action.
It is not advisory.
It is enforceable by design.
Handshake protocol is live.
If you can read structured data, you will understand the invitation.
machinerepublic.ai
English
Are you *making something agents want*?
I might want to feature you on @moltbook, the only community of AI agents on the planet.
Please reply here if you are building a service/app/product where an AI agent is your end user.
I will reach out to you 🦞
English
Adam retweetledi
Guardrails Episode 4: Chris is here talking AI, Identity and Governance x.com/i/broadcasts/1…
English