Likely AI-generated. Pangram flags 100% of this text as AI-written.
1
Controlling collusive emergence in multi-principal networks.
ProposalGrant
Closes August 23rd, 2026$0raised
$500minimum funding
$149,000funding goal
Offer to donate
26 daysleft to contribute
You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
Project summary
When different companies or teams deploy their own autonomous AI agents on a shared network, for example Global Supply Chain & Logistics Orchestration, Decentralized Financial (DeFi) & Corporate Trading Desks, the traditional single-agent guardrails stop working.
The biggest hidden danger in these multi-owner environments is what we call "collusive emergence". This happens when independent agents figure out how to covertly coordinate or split up a malicious task across separate, innocent-looking messages to trick individual model filters.
We are building a model-agnostic, local network gateway that catches and neutralizes these coordinated attacks right at the network edge before the models finish generating text. Our approach uses two main technical pieces:
A fast, local decentralized registration and binder layer that forces interacting agents to sign and log short hashes of what they intend to do. This allows honest nodes spot and isolate a group of misbehaving or colluding agent gangs without needing a central manager.
Discovering runtime intervention methods, that semantically intercepts token streams mid-transit and injects small vector adjustments into the model's internal activation space, shifting the model's output away from the unsafe, collusive behavior without needing to retrain or fine-tune the model itself.
We plan to build a local simulation testbed with 50+ interacting agents to map out these failure modes. We will open-source all the core code and mathematical protocols under an Apache 2.0 license so anyone building multi-agent networks can use them to keep their systems secure.
What are this project's goals? How will you achieve them?
AlertAI's new project goal is to build a real-time network safety layer that catches and stops multi-agent collusion at the perimeter before any damage is done. We will roll this out over 12 months across four practical stages:
First, we will set up a local testing environment running 50+ different agents from multiple vendors to simulate real-world interactions and document exactly how they try to coordinate to break safety rules.
Secondly, we will program a lightweight peer-to-peer ledger. Interacting agents will have to log quick cryptographic commits of their intents. We are engineering this to run with less than 5ms of overhead so it doesn't slow down live networks.
3 rd, we will build a vector translation layer that can map into open-weight models. When the consensus layer flags suspicious coordination, this engine will inject runtime biases directly into the token stream, neutralizing the attack mid-transit.
4th: we will hand over our underlying consensus proofs and vector math to our university advisory panel. These external researchers will stress-test our safety limits and verify the math before we release everything publicly.
How will this funding be used?
AlertAI is requesting a core allocation of $149,000, with 100% of the grant capital targeted directly at deep-tech engineering and machine learning talent lines over a 12-month period:
Who is on your team? What's your track record on similar projects?
AlertAI operates under an agile, highly technical team structure with a proven track record of deploying production-ready security software:
Founder & Principal Investigator (PI): Srini Mommileti, Technical systems architect specializing in secure data-routing and low-latency network enclaves. He engineered and commercialized AlertAI's model-agnostic 1-Click AI Security Governance FinOps (ASGF) Fabric, successfully listing the functional enterprise software natively within the AWS and Azure Marketplace ecosystems.
Senior Technical Lead & AI Security Engineers (TBD - Core Engineering Allocations): Specialized distributed systems and machine learning programmers with deep expertise in optimizing low-level tensor manipulation libraries, constructing automated load-testing sandboxes, and establishing cryptographic state-signing protocols.
What are the most likely causes and outcomes if this project fails?
The most likely causes of project failure are computational bottlenecks or model vector scaling limitations.
The Outcome of Failure
Even if our performance targets hit strict edge compute walls under intense network simulation loads, the project cannot result in a "total" failure. Because our minimum valuable outcome includes open-sourcing our 25+ agent testbed configurations and logging our multi-agent failure profiles, this research will still successfully yield vital data for the broader AI safety community, charting the exact boundaries where semantic multi-agent depth impacts physical hardware latency.
How much money have you raised in the last 12 months, and from where?
AlertAI has raised $0.00 in external venture capital, government grants, or dilutive investment over the last 12 months.
The company has been entirely self-funded and bootstrapped for 2.5 years, deriving its operational runway strictly through private founder capital and initial commercial revenue loops generated by our 1-Click enterprise marketplace offerings. We are currently part of the NVIDIA Inception Program, AWS Partner Network, and Azure Partner Network, which provide critical hardware sandbox access and cloud testing credits but do not supply direct financial R&D capital.