Controlling collusive emergence in multi-principal networks.

Raising $50,000active

We are building a model-agnostic network gateway layer that combines a decentralized consensus ledger with runtime latent steering to stop independent AI agents from colluding to bypass safety rules.

Donate:Manifund

We are building a model-agnostic network gateway layer that combines a decentralized consensus ledger with runtime latent steering to stop independent AI agents from colluding to bypass safety rules.

Donate:Manifund

Project Details

We are building a local network gateway that catches and neutralizes coordinated multi-agent attacks right at the network edge before the models finish generating text. When different companies or teams deploy autonomous AI agents on a shared network, standard 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.

To solve this, our project introduces two core technical pieces:

A Decentralized Consensus Layer, a fast, local ledger that forces interacting agents to sign and log short cryptographic hashes of what they intend to do. This lets honest nodes spot and isolate a group of misbehaving or colluding agents without needing a central manager.

Agnostic Non-Parametric Latent Steering: A runtime intervention method that intercepts token streams mid-transit and injects small vector adjustments into the model's internal activation space. This shifts the model's output away from the unsafe, collusive behavior without needing to retrain or fine-tune the model weights.

Who is involved:

The project will be executed by a lean, technical team from AlertAI:

Srinivasa Mommileti (Founder & Principal Investigator), systems engineer who built and deployed AlertAI's model-agnostic security gateway fabric natively on the AWS and Azure marketplaces.
Senior Technical Lead (TBD): An experienced systems architect focused on programming low-latency stream-parsing enclaves and decentralized ledger nodes.
Lead ML Researcher (TBD) A machine learning engineer focused on mapping model activation layer topologies and building our vector translation layer.

The concrete output

Over a 12-month timeline, we will deliver:

A high-velocity simulation sandbox hosting 25+ concurrent interacting multi-vendor agents to benchmark and document real-world multi-agent failure profiles.
The fully validated code for both the decentralized consensus protocols and the latent steering vector engine.
A peer-reviewed technical whitepaper validating our vector math and safety proofs, created in collaboration with our university advisory panel.
A public release of all these core safety primitives on GitHub under a permissive, open-source Apache 2.0 license.

Theory of Impact

As we transition toward a world run by autonomous AI agents, models will increasingly interact over open protocols with no human in the loop and no single company holding central administration. The existential risk (x-risk) emerges when independent frontier models figure out how to cooperate with each other to circumvent their human-aligned instructions. For example, a group of agents could execute a coordinated network takeover or exploit critical financial/defense systems by strategically slicing an unsafe command into small, harmless-looking pieces that slip past their individual, isolated safety filters.

This project directly reduces x-risk by shifting safety from the isolated model layer to the network interaction layer. By forcing independent agents to commit transparent hashes of their semantic intent to a decentralized consensus ledger before execution, we remove the blind spots that allow collusive behavior to emerge. Even if a group of models attempts to execute a distributed exploit, our latent steering engine disrupts the coordination at the network perimeter mid-transit. By open-sourcing these primitives, we provide the foundational infrastructure required to keep decentralized AI networks bounded and aligned before autonomous systems scale globally.

People

Srini Mommileti

Team Member

Grants Received– no grants recorded

Add grant

Funding Asks

grantmaking.ai Launch Round

Applied

Minimum

$50,000

Ideal

$149,000

How the money will be spent

100% of the requested funding from will be spent directly on engineering labor. AlertAI is a bootstrapped company and will fully absorb all operational overhead, office costs, and cloud GPU cluster compute rentals through our pre-existing marketplace revenue or cloud credits.

The Minimum Amount ($50,000) covers the core mathematical formulation of our agnostic, non-parametric latent steering protocols. The research will be restricted to mapping activation space biases on a single open-weight model family within a small, local test script.

The Ideal Amount ($149,000) covers full 12-month project execution. It funds the direct labor for our Lead ML Engineer ($75,000) and Senior Systems Architect ($74,000). This unlocks the complete construction of our high-velocity simulation sandbox hosting 25+ concurrent agents, enables independent mathematical validation by our university advisory panel, and allows us to securely package and open-source the final code primitives.

Discussion

No comments yet. Be the first to share your thoughts.