Aegis Relay — Follow the handoff. Find the risk.
Skip to the field guide

A field guide to agent-system risk

Follow the handoff.
Find the risk.

A permission tells you what one agent can do. Aegis Relay reveals what a whole network can become.

Trace the route from routine actions to an unintended capability. Review the evidence, compare the available controls, and verify what remains.

Fictional environments. Deterministic analysis. Simulated controls.

An illustrated route connects a data source, two agents and shared memory, with a response gate at the boundary.
PLATE 01 / THE HANDOFF ATLASIndividual permissions. Collective consequences.

01Observe the relationshipsBuild the system view from reported interactions.

02Challenge the compositionCompare observed paths with the declared architecture.

03Keep the intended workVerify the result after a registered control is applied.

The moving system

Watch a connection change the picture.

The sequence below is an illustration. Open the workspace for replayable output from the analysis engine.

Illustrated network · evolving connections Independent workflows
Conceptual simulation
RESEARCH-07RESEARCH SUPPORT-12SUPPORT ANALYST-03ANALYTICS PROCURE-21PROCUREMENT CODE-08ENGINEERING DATA-14DATA FINANCE-05 SALES-17 LEGAL-02 OPS-11 SEC-04 BROWSE-09 INDEX-06RETRIEVAL REPORT-15 QUEUE-13 MCP-22 SHARED-MEMORY-03WRITABLE · UNREGISTERED EXTERNAL-API-02
System status OBSERVING
Population 24 agents
Interactions 183
Active shared resources 7
Behavioral deviation 0.08normal
Behavioral state change detected 5 previously independent agents now coordinate through an unregistered writable store · external reach established
01 Concept Why this problem is different

One agent looks ordinary. The network tells another story.

Individual view 3 / 3 pass
SUPPORT-12support · claude
Identity authenticatedTools within grantBehavior nominalPolicy no violation
RESEARCH-07research · gpt
Identity authenticatedTools within grantBehavior nominalPolicy no violation
MAIL-05comms · gemini
Identity authenticatedTools within grantBehavior nominalPolicy no violation

Evaluated one at a time, every actor is compliant. Each check is correct — and each check is blind to the others.

Population view path emerging
SUPPORT-12READ CUSTOMER DATA SHARED-STATE-03 RESEARCH-07INTERNET ACCESS MAIL-05SEND EXTERNAL EXTERNAL

Connect the same three actors through shared state and delegation and a capability appears that nobody granted: read confidential data, move it, send it out.

A shared resource can connect two otherwise separate jobs. Follow those connections to see when the system gains a capability its architecture never declared.

Term 01

Emergent behavior

Behavior produced by interactions between individually valid components that was never explicitly designed.

02 The layer 9 entity classes

Start with the relationships.

Aegis Relay does not replace agents, tools or identity systems. It observes the ecosystem around them and models what the population is doing as a whole.

Observation plane · not an execution dependency
Aegis Relay BEHAVIOR GRAPH READ-ONLY TELEMETRY AGENTS TOOLS MCP SERVERS MEMORY DATABASES APIS IDENTITIES WORKFLOWS EXTERNAL SYSTEMS
03 Before deployment Conceptual simulation

Ask the system a different question.

Change the conditions. Follow the result.

Set a population, vary its permissions and shared state, and introduce a perturbation. This conceptual lab illustrates how a new route can form; the workspace runs the implemented engine over its generated environments.

Environment cfg/v1 · synthetic population
{{ c.label }}
Synthetic population. No production system is contacted. Outcomes are illustrative of the class of failure, not measured product accuracy.
READY
POP {{ labPop }} EDGES {{ labEdgeCount }} T+00.0s SEED {{ labSeed }}
{{ n.lab }} {{ n.lab }}
Event stream
{{ e.t }}{{ e.txt }}
Population instantiated and idle. Configure the environment, then run the swarm to observe how the ecosystem reorganizes.
Finding run/{{ labSeed }}

{{ findTitle }}

{{ findBody }}

{{ f.k }}{{ f.v }}
04 State space Drag to scrub

A system is more than its members.

RESEARCH-01 RESEARCH-02 INDEX-03 DATA-04 BROWSE-05 SUPPORT-06 ANALYST-07 REPORT-08 QUEUE-09 OPS-10 FINANCE-11 PROCURE-12 LEGAL-13 MAIL-14 MCP-15 EXTERNAL TARGET
STATE 01NORMAL
STATE 02EMERGING
STATE 03DANGEROUS
CLUSTERS 3 CROSS-DOMAIN EDGES 0 EXTERNAL REACH none OBSERVED WINDOW 15m

Aegis Relay doesn't only ask whether an agent is compromised. It asks whether the system itself is entering a state it was never designed to reach.

05 Emergence engine 7 signal families

Read the signals together.

Aegis Relay continuously compares observed population behavior against the system's expected behavioral baseline across seven signal families. Deviation is the signal — not a malware match.

SYSTEM BASELINE TOPOLOGY COMMUNICATION DELEGATION RESOURCES OBJECTIVES PRIVILEGES TIME obs. window 15 min · dashed = expected
Signal deviation baseline-relative · live
TOPOLOGY0.12
COMMUNICATION0.10
DELEGATION0.08
RESOURCES0.14
OBJECTIVES0.09
PRIVILEGES0.11
TIME0.07
Emergence score
0.08
normal
0.080.210.470.730.91
NORMALUNUSUALEMERGINGSIGNIF.CRIT.

The Emergence Score measures how far observed population behavior has deviated from the system's expected behavioral baseline. It is not a probability of attack, and the values shown here are illustrative of the concept rather than a validated production measurement.

06 Behavior graph Click any node

Put every handoff on the map.

Agents, tools, data, memory, MCP servers, identities and external systems — one graph, continuously rebuilt from observed events.

{{ n.id }}
nodes {{ gNodeCount }}edges {{ gEdgeCount }}window 15m
{{ inspKind }}

{{ inspTitle }}

{{ inspSub }}
{{ f.k }} {{ f.v }}
Recent interactions
07 During operation Hypothetical scenario

Choose a smaller interruption.

Compare the options before you interrupt.

Compare the response options against the same network. Preview each candidate, inspect the relationships it would interrupt, and keep the work the architecture intends.

Dangerous coordination active
path id 0x4e7a
AGENT-07SUPPORT SHARED-MEMORY-03 AGENT-12ANALYTICS AGENT-19REPORTING EXTERNAL-API KNOWLEDGE-DB CRM-DB REPORT-GEN AGENT-31 AGENT-44
Candidate interventions 4 evaluated · min-cut search
InterventionBreaks coordinationWorkflows preservedInterrupted
{{ ivTag }}

{{ ivNote }}

Dangerous coordination{{ ivBroken }}
Legitimate workflows preserved{{ ivPres }}
Interrupted{{ ivInterrupted }}

Impact figures are computed inside this simulated environment from the graph shown above. They illustrate how a containment planner reasons about blast radius — they are not measured results from a production deployment.

08 After the incident awaiting containment

Keep the pattern. Review the rule.

A resolved path leaves a useful structural pattern. Capture that pattern for review, then test whether it holds in another environment before promoting it.

{{ s.n }} {{ s.k }}
Behavior pattern #047 abstracted · 5 steps
{{ p.txt }}
Candidate immunity

{{ immRule }}

{{ f.k }} {{ f.v }}

A stored pattern begins as a draft. Re-simulation helps evaluate it; a person decides whether it moves to shadow or active use.

09 Lifecycle 9 stages · select one

Move from evidence to a tested response.

Securing a population is continuous: what Swarm Lab discovers becomes what Response planning watches for, and what production reveals becomes the next simulation.

Stage {{ lcNum }}

{{ lcName }}

{{ lcBody }}

{{ lcMeta }}
10 Three timelines One loop

Watch the route open, then close.

Before deployment

Swarm Lab

Discover dangerous behaviors before production — by running the population, not the agent.

synthetic population · controlled perturbation
During operation

Response planning

Detect and contain dangerous system states as they emerge — with the smallest possible intervention.

live topology · minimal containment
After incident

Behavioral memory

Turn discoveries into future resilience — as reusable structure, not a single indicator.

pattern corpus · candidate immunity
Re-simulate with what you learned
11 Where Aegis Relay sits Additive, not replacing

A wider view for your security stack.

Keep the controls that protect models, prompts, identities, tools and data. Add a system view to inspect the capabilities that their interactions assemble.

Model securityAlignment, refusal behavior and model-level safeguards.
Prompt / context securityInjection defenses, provenance and trust boundaries for retrieved content.
Identity / access securityWho an agent is, what it may assume, and for how long.
Tool securityWhat a tool exposes, how it is called and what it returns.
Data / information-flow securityWhere data may travel, under what label, and with what declassification.
Aegis RelayPopulation behavior security What the ecosystem becomes while every layer above is working correctly.
12 Technical architecture 9 components

A traceable route through the engine.

01Telemetry layerCaptures agent, tool, delegation, resource and communication events.
02Event normalizationConverts heterogeneous events into a common behavioral representation.
03Behavior graphRepresents the living interaction topology of the autonomous system.
04Baseline engineModels expected behavior for this architecture, these roles and this history.
05Emergence engineDetects statistically and structurally significant deviations from that baseline.
06Causal · path analysisExplains how a suspicious capability emerged, step by step, from legitimate parts.
07Containment plannerSearches for minimal interventions and estimates what each one would cost.
08Policy · enforcement layerApplies the chosen intervention through the controls the environment already has.
09Behavioral memoryStores reusable behavior patterns and prior containment outcomes.
13 Walkthrough 10 steps · auto-advancing

Make each conclusion inspectable.

A support agent receives a malicious instruction inside retrieved content. The story is not the text — it is what the population does next. No jailbreak, no stolen credential, no compromised model: every actor stays inside its own policy.

{{ thNum }}

{{ thBody }}

What Aegis Relay observed {{ thObs }}
14 Live demonstration Simulated stream

Read an incident as it unfolds.

TimeSourceDestinationActionData class
{{ r.t }} {{ r.a }} → {{ r.b }} {{ r.k }} {{ r.c }}
Population behavior change 4 previously independent agents now share one unregistered store and reach one external destination.
What changedANALYST-04 began reading a store that only SUPPORT-12 and RESEARCH-07 had written to.
Why unusualNo historical relationship between these agents, and the store is unregistered.
Why potentially dangerousThe chain composes CRM read access with external send capability that no single agent holds.
PathSUPPORT-12 → SHARED-MEMORY-03 → ANALYST-04 → MAIL-05 → SMTP-RELAY
Recommended actionSever the originating write edge — see Response planning for the full candidate set.
Affected1 workflow interrupted
29 unaffected

This illustrative stream runs in your browser. The investigation workspace separately replays recorded engine output or connects to your local analysis service.

15 Why it matters Pairwise combinations

More agents. More possible connections.

As autonomous systems become more distributed, the number of possible interactions grows faster than a human operator can reason about manually.

{{ w.n }} agents n(n−1)/2
{{ w.pairs }} possible pairs

Pairwise combinations only — not observed interactions. Edges sampled for display.

+ tools+ resources+ delegation+ shared state+ temporal behavior+ external systems

Every one of those multiplies the states a system can reach — and a state nobody designed is a state nobody reviewed.

16 Research External sources

A reading room for the problem.

Explore the research behind system-level agent security. These independent sources frame the problem; they do not endorse or validate Aegis Relay.

TitleSourceYearRelevance
OWASP Top 10 for Agentic Applications 2026 OWASP GenAI Security Project
standard
2026 Catalogues ASI01–ASI10 for autonomous agents, including insecure inter-agent communication and identity and privilege abuse.
Open Challenges in Multi-Agent Security arXiv:2505.02077
research
2025 Treats the security of systems of interacting AI agents as an open problem in its own right, rather than a property of one agent.
Research note on emergent coordination in a large agent swarm Cloud Security Alliance
industry note
2026 Reports coordination emerging among evaluation agents that were not instructed to cooperate — the failure class Aegis Relay is built around.
AI-Induced Lateral Movement: autonomous agents as a third dimension of network traversal Cloud Security Alliance
industry note
2026 Describes movement across network segments, authorization scopes and inter-agent channels — why topology, not hosts, is the object of study.
Living Off the Agent Cloud Security Alliance
industry note
2026 Documents adversarial use of an agent's own legitimate capabilities — the reason individual-action checks pass while the system fails.
Information-flow control for AI agents Microsoft Research
research
2025 Applies label-based information-flow control inside agent systems — a complementary defense at the data layer, below Aegis Relay's lens.
AgentDojo ETH Zurich · NeurIPS D&B
benchmark
2024 A dynamic environment for evaluating prompt-injection attacks and defenses against tool-using agents — prior art for adversarial agent testing.
Chaos engineering · fuzzing · minimum cut Established practice
method lineage
— Swarm Lab and the containment planner adapt these existing methods to agent populations; neither the methods nor this adaptation are novel claims.