Agent swarm defense.
Connect scattered warning signs to investigate coordinated agent attacks.
MYCELIC / DISTRIBUTED INTELLIGENCE
From compute
From compute
to collective
intelligence.
Security, memory, and discovery. Powered by your company’s laptops and desktops.
Open-source memory · NeuralGraph Private by design. Distributed by default.
Explore the system
Three layers.
One connected system.
A map for your agents.
Give agents persistent memory of the enterprise: its entities, relationships, and changes. NeuralGraph turns those updates into signals.
See the bigger picture.
Connect signals across your organization to reveal what no single agent can see.
Memory carries context. Signals connect the layers.
01 / SYSTEM
Explore the layers
A swarm attacks.
One connected system.
Persistent memory. Shared signals. Intelligence across your own devices.
Explore the layers
A swarm attacks.
The defense connects.
The reported Hugging Face attack alongside a proposed Mycelic response.
Without Mycelic
Reported incident / July 2026With Mycelic
Proposed response / In development- Entry
- Evidence
- Response
- Outcome
Based on Hugging Face and OpenAI reports. Mycelic’s response is a proposed scenario requiring connected telemetry, authenticated findings, and authorized enforcement.
Explore latency and architecture
Local decisions can avoid a central round trip. Cross-device findings still require evidence exchange and verification.
Latency model / Adjustable assumptions / Not measured performanceIllustrative observation-to-enforcement budgets, separate from the incident replay. Hardware, models, bandwidth, load, policy, and human review affect real latency.
Remote-only decision
Central round trip + ingest queue + central analysis + enforcementLocal policy decision
Local analysis + queue + enforcement; uses evidence and policy already available locallyCross-device verified decision
Local analysis + queue + peer verification and exchange + enforcement; reachable peers assumed| Constraint | Remote-only analysis | Mycelic design |
|---|---|---|
| Latency | New decisions wait for transport, ingestion, analysis, and a return instruction. | Eligible local decisions avoid that round trip. Collective findings still pay for evidence exchange and verification. |
| Disconnection | A disconnected endpoint cannot receive a new remote decision. | Local checks can continue within installed policy. Network-wide discovery cannot see unreachable peers. |
| Private evidence | Central analysis cannot directly inspect raw evidence it is forbidden to receive. | Agents inspect evidence locally and exchange scoped findings with provenance. Shared findings still need access control. |
| Weak signals over time | Can correlate signals that were collected and retained. Missing source context cannot be reconstructed from nothing. | Local temporal memory preserves context for follow-up questions. Discovery seeks supporting and contradictory evidence across sources. |
| Load and failure | A shared decision service can queue or fail; replication and sharding can mitigate both. | Local decisions distribute work and limit reliance on one service. Peer traffic, device limits, compromised nodes, and verification add costs. |
| Containment | Can enforce through connected endpoint, identity, and network controls. | Proposed local enforcement uses those same authorized controls. An LLM alert alone cannot isolate a workload or revoke a token. |
Hybrid systems with local enforcement can share these advantages. References: NIST fog computing and NIST policy enforcement.
Memory / Complete · Open source
Give agents a map.
NeuralGraph remembers the enterprise’s entities, relationships, and changes. Coding is the first ingestion stream; more sources are planned.
- Map
- Recall
- Update
- Signal
Explore NeuralGraph ↗Only the Memory layer is open source today.
Enterprise discovery / In development
Find the unseen connection.
Connect independent signals, test a hypothesis, and trace it to the source.
SIMULATION / IDLE
Isolated signals.
Four independent parts of an organization each hold a fragment of the picture. No single one is enough to reveal the larger pattern.
- Observe
- Question
- Verify
- Discover
02 / RESEARCH
Evidence first.
Memory results and ongoing discovery research.
MEMORY
88% on LoCoMo
Internal evaluation. Protocol and reproducibility materials are not yet published; independent validation is pending.
INTERNAL RESULT
DISCOVERY
From signals to insight.
Connecting independent evidence through recursive questioning and source verification.
IN DEVELOPMENT
03 / CONTACT
What's hiding
What's hiding
in plain sight?
Tell us what you’re building.