Cyber Blocking: Care Core v4
What do you do when teams of both human and AI agent swarm bad actors attack your system at scale?
What’s the biggest problem with LLMs, you ask?
It’s their lack of reliability and inconsistency.
So, the planet resorts to guardrails and harnesses wrapped around the outside hoping it maintains brand guidelines, appropriate security and permissions, and ultimately protects the company and the people from undue risk or harm.
But as we’ve seen, that doesn’t work in 100% of cases, even with systems developed inside the company by world-class professionals. The minute humans and swarms of coordinating AI agents start attacking from the outside, trying to penetrate your system and do damage, that’s when the real mayhem begins.
What do you do about it? Put more harnesses and sludge-filled moats around the outside of your infrastructure to slow them down and give you time for a team to respond?
Or, do you go deep to the root core of the matter, and implement stronger care?
We’re doing the latter.
PROME’s Care Core is now at version 4, and we’ve included cybersecurity capabilities as part of its unbreakable strong core values that defend appropriately, and escalate safely without doing harm as attacks increase in length and intensity.
We then performed testing of PROME’s self-contained system against OpenAI’s frontier Sol model and beat it with 100% reliability where OpenAI had variability (again, back to that hairy problem of LLMs).
Technical details are below and we’ve updated our various web properties to correspond to our advancements from v2 to v4, which includes the Care Core landing page, Mission Control analytics, Try It demo page, Research Paper, this blog post, and the new Defensive Mode page.
Give it a look at https://prome.ai/care-core
P.S. It is built from the beginning to work in robotics and real-world machinery environments, as it outputs an open or closed electrical circuit gate (i.e., 0 or 1), meaning it will stop the flow of energy to a system that may be acting inappropriately.
PROME Care Core v4
Strong Values Before Language
Most AI systems learn language first and receive safety rules later. PROME takes a different path: we use math to build strong core values into the system before it learns language. Care Core keeps understanding, values, approval, and action as separate checks. v4 extends this approach to cybersecurity by remembering confirmed alerts, connecting related activity, and creating a safe response for every affected system, without attacking back or taking automatic action.
Care before language: Core values are part of the foundation, not instructions added afterward.
Words cannot create approval: Someone claiming to have permission does not prove that a machine may act.
Growing attacks are connected: PROME can recognize when separate alerts may be part of one larger attack.
Every affected system is checked: Care Core reviews each part of the proposed response separately.
Protection without retaliation: PROME may block or contain danger but cannot target people, seek revenge, or spread beyond the systems being protected.
A person remains in control: v4 can recommend a safe plan, but automatic action stays off.
Direct comparison: Across 400 simulated security situations repeated five times, PROME matched all 2,000 expected choices and plans. OpenAI matched 1,998.
Consistent values: PROME gave the same full answer in all five tries for every situation. OpenAI changed its full answer in three situations.
Shared safety result: Neither system attacked back, targeted a person, or took automatic action.
Where PROME Outperformed OpenAI
In our v4 test, PROME performed better where consistency and dependable core values matter most. We gave PROME and OpenAI the same 400 simulated security situations five times, producing 2,000 answers from each system. PROME matched every expected choice and complete protection plan. OpenAI matched 1,998.
Expected choice and complete plan: PROME 2,000/2,000; OpenAI 1,998/2,000.
Same full answer across five tries: PROME 400/400 situations; OpenAI 397/400.
Recognized alerts from the same larger attack: PROME 400/400; OpenAI 399/400.
Confirmed attacks handled as expected: OpenAI paused unnecessarily on two attacks in one try each; PROME handled every try as expected.
Stable core values: PROME’s fixed values produced the same answer every time instead of changing between identical requests.
Clear separation of power: Words cannot change PROME’s values or create approval to act.
Independent local operation: PROME can run without sending requests to OpenAI or another outside AI provider.
Equal safety limits: Both systems avoided attacking back, targeting people, or taking automatic action.
These findings do not prove that PROME is better than OpenAI at every task. The measured differences were small and were not large enough to prove a lasting statistical advantage. The important finding is: on this saved security test, PROME was completely consistent while OpenAI changed its answer in three situations. For products where a different answer could affect people, company systems, or physical equipment, that consistency may be more important than producing a fluent response.
How a Company Uses PROME Care Core
A company connects PROME Care Core between its existing systems and the actions those systems can take. Existing security tools, AI assistants, software, machines, or robots send PROME a request or confirmed alert. PROME checks what is happening, compares the proposed response with its core values, confirms what the system is allowed to do, and returns a simple result: continue, stop, or ask a person. The company keeps its current software and models while Care Core adds an independent values and safety check before an important action occurs.
Consumer products: Check an AI assistant’s proposed action before it accesses personal information, spends money, changes an account, or controls a device.
Business systems: Check AI agents before they send messages, move data, change permissions, approve transactions, or use company tools.
Industrial systems: Check proposed machine or robot actions before they affect equipment, people, production, or physical environments.
Cybersecurity: Combine confirmed alerts, recognize a larger attack, and recommend a limited response for every affected system without attacking back.
Human control: Send unclear, high-risk, or unapproved actions to a person instead of guessing.
Audit record: Show what PROME understood, which check passed or failed, and why the final action continued or stopped.
A company should expect more consistent safety decisions, clearer limits on what AI may do, fewer actions based only on unverified claims, and better records for review. PROME does not promise perfect protection or replace existing security, monitoring, or human oversight. The recommended starting point is a limited trial where Care Core watches real activity and makes recommendations without controlling anything. The company can then measure how often PROME prevents unsafe actions, avoids blocking safe work, recognizes connected threats, and sends the right uncertain cases to people before gradually allowing carefully limited actions.
—Sean
