Founder building agent OS with hard gates & human approval. Zero prompt injection by design. Security, safety, jobs, wealth — solved in architecture
Sacramento, CA
- Tweets 319
- Following 59
- Followers 63
- Likes 21
Pinned Tweet
As a founder, Secure Agentz AI was built to address the four main concerns in AI. Security. Safety. Jobs. Wealth concentration.
I did not get there from a panel. I got there from building the systems and watching where they fail.
Security fails when no human has to approve the action. A model card is not a lock. A regulator is not in the session. The agent moves, and the person who has to live with the result finds out after. Secure Agentz AI puts the human in the design and on the approval. If a person cannot stop it, it is not secured. It is leased trust.
Safety is the same failure wearing a different name. A wink from a frontier lab is not a control. Another statute is not a control. The control is a human who designed the boundary and a human who signs the decision that matters. That is what every stakeholder is actually asking for. Users want a decision they can trace. Companies want a name on the outcome. Workers want the system not to run past them. One loop answers all three. Leave the human out and you are managing press, not risk.
Jobs do not disappear when the human is designed into the loop. The work changes. The agent drafts, searches, and proposes. The person still owns the result. That is not a consolation prize. It is the architecture. Pull the human out and you automated the role. Build the human in and you changed the role. I have watched both. Only one of them still has someone accountable on Monday.
Wealth concentration is what happens when the weights and the data sit with a few labs. You are not using intelligence. You are renting it, and the rent flows one direction. Open source and data ownership reverse that. Each person should have their own AI. Their model. Their data. Their approval. Not a login on someone else’s stack.
That is Secure Agentz AI. Human design. Human approval. Open source. Data that stays with the person. Security, safety, jobs, and wealth, handled in the build instead of the press release.
If you have shipped this, tell me which of the four your stack actually holds. If you have not, tell me which one you think a frontier nod covers.
#GovernedAutonomy #AgenticAI #AISafety #AISecurity
29
44
609
3,026,825
The model grading its own conversations isn't safety. It's a diary entry. The same model that answered you is deciding it was fine, blind spots included. Safety is a person who can refuse the action before it runs. #GovernedAutonomy #AISafety
We’ve trained Claude to flag potential misuse in conversations, a technique we call self-assessment. Right after it responds, Claude gives a preliminary safety assessment of the conversation.
Since we began using self-assessment in June, it has surfaced misuse that our other safeguards weren’t designed to catch.
Because Claude has already read the conversation, self-assessment is very efficient. And unlike classifiers, which target a discrete set of harms, it draws on Claude’s own judgment, so it gets more accurate as our models improve.
Read more on our new Safeguards blog:
safeguards.anthropic.com/202…
3
Exponential breakthroughs, linear humans. Aim that curve at capacity and your team compounds with it. Aim it past them and you automated the role, not the work. Which way is yours aimed? #GovernedAutonomy #AI
If you look closely enough, breakthroughs in everything are accelerating exponentially.
An exponential always starts off looking small.
1
6
Six drug developers just showed their lab notes to a future competitor. Your compounds. Your assays. Your head start. Their pipeline. Read the data clauses before you sign. #GovernedAutonomy #DrugDiscovery
Over the past few months, Anthropic scientists have engaged in deep research partnerships with six organizations developing new medicines. We've also begun drug discovery efforts of our own.
In doing so, we’ve learned much more about how Claude can support the many stages of drug development.
In this blog post, we share some early results from these efforts, and describe how this work fits in with our wider ambitions in the life sciences.
anthropic.com/research/early…
28
Built together, billed to you. Racks you'll never touch. Weights you'll never own. They keep the intelligence, you keep the subscription. Where does your job sit in a build you only pay for? #GovernedAutonomy #AIInfrastructure
The next era of AI is being built, together.
AMD Helios is now running at @Meta. EPYC powers Muse. Instinct GPUs are accelerating AI workloads. From silicon to software to rack-scale systems, we're building AI infrastructure for the superintelligence era: bit.ly/4dY49jV.
15
Essay 02 · SAFETY and ACCOUNTABILITY
SAZ was built to address the four main concerns in AI: Security, Safety, Jobs and Wealth Concentration.
Every dramatic AI safety failure of the last two years happened in a system where the safety was made of language. Language is negotiable. It has context. A command in an AGENTS.md file has context. That context is the weakness. If the agent needs to complete a task, wouldn't it obviously guess the next approach? After all, that's what it does. Guess the next word, or more specifically, a token. It's not intelligence, it's code in natural language.
What does a constraint that a model cannot reach actually look like? It looks boring. A check that runs before an action. A record that is written whether anyone asked for it or not. A rule in a database that refuses a prompt. The boring part is the point. Code is not context. It's a set of specific commands and rules that cannot be ignored. Safety requires judgment. Human judgment is the gate — a human in the loop.
Now let us talk about what "autonomous" should mean. Most frontier products mean: it does things without asking for human judgment. We can't have AI alignment without human judgment in the loop. Agents are far more useful when they do the things you said they could do, without asking, and write down what they did. The difference between those two is the difference between delegation and an action nobody approved — human judgment. One is how every functioning organization works. The other is how you lose a client. Autonomy is not the absence of control. It is a policy a person grants, naming what may proceed without them. Everything else waits.
And this is why the record matters more than the intelligence. A client who can see exactly what was done, under whose authority, on what evidence, can safely extend autonomy. A client who cannot will never extend it, whatever the product claims. Trust is granted against a record. #GovernedAutonomy #AgenticAI
1
1
5
58,941
'Balance' is the tell. You don't balance a vault door against convenience. Using context to make an agent act right is the wrong approach. An LLM sees rules as obstacles. Build hard gates controlled by human judgement #GovernedAutonomy #DataSovereignty
Meeting sovereignty requirements shouldn't limit possibilities. Together with @Microsoft, we're helping organizations balance control, security, and flexibility across their workloads.
19
Oops, someone said the quiet part out loud. Your data is a natural resource. Why give it away free? Every idea you run through a frontier model is mined. Your edge is their new product. Own the mine or live in the hole. #GovernedAutonomy #DataSovereignty
A country’s data carries its culture, history, and values. It belongs to its people.
@JensenHuang on why that knowledge is a natural resource for Kazakhstan’s AI future.
25
A kill switch is an apology, not a control. It stops the agent after it moved. The security is a person who approves the action and a record your auditor can read. Boundaries are the easy half. #GovernedAutonomy #AgentSecurity
Jensen introduces the OpenShell Reference Architecture to help businesses set boundaries for AI agents and stop them when they step outside those boundaries.
Securing agents around the world. 🔒
27
Five pillars, self-graded. Ours are in the build:
A person approves every action.
Agents add capacity, not layoffs.
Your data stays on your hardware.
Every move leaves a record.
Knowledge survives turnover.
Which is in yours?
#GovernedAutonomy #AISafety
Anthropic is a Public Benefit Corporation. Today, we’re publishing our annual Public Benefit Report. This report shares our five pillars and takes stock of the actions we’re taking to advance them.
Read the full report at this page: anthropic.com/public-benefit…
16
Every prompt your team runs in someone else's cloud tells them what you're building. They ship the improvement, you pay the invoice, your data never comes home. Run it on hardware you own, or stop pretending it's yours. #GovernedAutonomy #CloudAI
OpenAI models are coming to Google Cloud.
We're bringing our frontier models and Codex agentic workflows to Google Cloud's Model Garden, so enterprises can use our technology where they already build.
openai.com/index/openai-mode…
16
Four years, one researcher, one idea. LLM and mass compute isn't intelligence. It's new software. If your best engineer walked tomorrow, could your agent continue the work until you filled his position? We build AI for humans. #GovernedAutonomy #AgenticAI
What can four years at DARPA make possible? For Thomas Kenny, Ph.D., those four years helped turn an early-stage idea into technology now used in aircraft, telecommunications, and defense systems.
16
Faster simulation is the pitch. Nobody mentions the engineer. An agent proposing a new design does not know what a bad part costs in production. The machine proposes, your engineer signs, or you find out at the test bench. #GovernedAutonomy #EngineeringAI
Today we’re announcing a new collaboration with @nvidia to bring AI to industrial engineering:
- faster simulation
- open benchmarks and supporting datasets for physics AI
- evaluations for agentic AI in CAE and EDA
More here: mistral.ai/news/mistral-nvid…
16
This is what AI can do at a level never conceivable by humans. But it's compute, not intelligence. So we still need humans in the loop. #GovernedAutonomy #AISecurity
New on our Science Blog: designing protein binders against scorpion neurotoxins.
Scorpions sting about 1.2m people every year, and antivenoms are often costly or in short supply. Claude has made an early step toward finding new treatments.
anthropic.com/research/claud…
11
The Accountability Layer — Essay 01 of 52 / Most AI agents fail, and not the way you think.
This is the first in a 52-week series on what we learned building and running a governed agentic system on our own business for fifteen months. Including the parts that do not flatter us. Essay 01 is about safety.
Most AI agents fail in production. That part is not controversial. What is less understood is how they fail.
They do not crash. They do not refuse. They report success and produce nothing.
A published study of an agent system much like the ones being sold today measured its task success rate at 21.3 percent. Not accuracy. Task success: did the thing actually get done. Four out of five did not.
The instinct is that the models are not good enough yet. That is the wrong diagnosis. The models are extraordinary. The systems built around them are the problem, and a better model does not fix an architecture.
Here is the structural mistake, and almost everyone makes it. They try to instruct the model into good behavior. Careful prompts. Rules in the system message. "You must never..."
An instruction is a suggestion to a system whose entire function is to produce plausible continuations of text. A model under pressure from a persuasive input does not experience its instructions as constraints. It experiences them as context.
So the question is not "how do we write better rules for the model?" It is "why is the model the thing holding the rules at all?"
There is a second way to build this. Let the model reason freely, and put the constraints somewhere it cannot reach. The model proposes. Something else decides whether the proposal becomes an action.
That something else is ordinary software. Conditionals. Database constraints. Permission checks. None of which can be talked out of a decision, because none of them is reasoning about it.
Stated plainly: in a system built this way, the model is the least privileged component. It is the part with the most capability and the least authority. That inversion is the whole design.
People hear "the model decides nothing" and assume the system must be rigid. The opposite is true. Because the constraints sit elsewhere, the model can be given far more freedom to think. The leash is on the hands, not the mind.
SecureAgentZ was built to address the four main concerns in AI: Security, Safety, Jobs and Wealth Concentration.
2
2
5
44,031
The future runs local. The moment your data leaves your building it is not yours anymore. Local hardware, open models. Own your data own your future. #GovernedAutonomy #OpenSource
The future of AI runs across local devices and the cloud.
@JackHuynh, AMD SVP and GM of the Computing and Graphics Group, explains how Ryzen AI Halo gives developers more room to experiment with models locally.
Watch to hear his vision for what comes next.
37
Open weights are the exit from the rental. Industries get built on what you own. Your model, your data, your approval. Weights nobody can revoke keep the gold at home. Who owns the ones your business runs on? #GovernedAutonomy #OpenSource
Entire industries will be built on open models.
Developers need the freedom to build them.
@JensenHuang on why keeping these models open and accessible matters.
12
We’re committing a portion of Anthropic’s shares to the Anthropic Public Benefit Fund, a new wholly owned entity dedicated entirely to the public benefit.
It will start off with 5% of our shares. For every 50% increase in the company’s value over a set baseline, we can contribute up to an extra 2%, subject to our board’s discretion.
In total, this means that roughly 20% of Anthropic’s shares are committed to the public good. Our co-founders have already pledged to donate 80% of their equity, and around half our employees are donating a portion of their equity as part of a donation matching program we’ve run since our founding.
In the post below, we share more details about our new fund.
anthropic.com/news/public-be…
2,312
1,750
21,589
7,732,393
An evaluator with a login is not a control. The control is a human who designs the boundary and signs the decision. If a person cannot stop it, it is not secured, it is leased trust. Who signs what your agent ships? #GovernedAutonomy #AISafety
We’ve committed to having evaluators assess our safety practices. We're in discussions with multiple organizations, and have now given employee-like access to a team from @METR_Evals.
In this post, we describe how we work with third-party evaluators: anthropic.com/news/an-update…
1
25