Discover
Register agents, resources, credentials and the dependencies between them: who delegates to whom, what each agent can reach, which workflows trigger which.
Agents & graphIncident response for interconnected autonomous AI
Your AI agents are connected. Your security isn't.
WHEN AN AGENT IS COMPROMISED _
CORVEN maps the dependencies behind autonomous systems, calculates the real blast radius of a compromised agent, executes minimum-safe containment and preserves verifiable incident evidence.
One compromised agent rarely stays one problem.
Agents hold delegated permissions, share APIs and wallets, and trigger each other's workflows. A stolen credential travels along those links. Revoking everything stops the attack and the business with it; revoking too little leaves an indirect path open. The hard part is knowing the smallest change that actually cuts the attacker off.
DELEGATION · SHARED RESOURCES · WORKFLOW TRIGGERSINDIRECT EXPOSURE
FIG. 01 — CONTAINMENT TOPOLOGYEXPOSED PATHS CUT · HEALTHY PATHS KEPT
How an incident is contained
01Dependency graph
02Minimum safe cut
03Verifiable evidence
CONTAINED
How CORVEN works
Discover. Analyse. Contain. Verify. Prove.
Each stage is a real engine behind the CORVEN API. The public demo runs all five on an isolated sandbox of five connected agents, with simulated executors standing in for the outside systems.
Run the demoFIG. 02 — INCIDENT RESPONSE LIFECYCLEDETERMINISTIC ANALYSIS · PLAN-BOUND APPROVAL · VERIFIED OUTCOME
- 01DiscoverTOPOLOGY · DEPENDENCIES
- 02TraverseBLAST RADIUS GRAPH
- 03PlanMINIMUM SAFE CUT
- 04ApprovePLAN-BOUND SIGNATURE
- 05EnforceGATEWAY · ON-CHAIN SEAL
Analyse
Deterministic graph traversal from the compromised agent: what is exposed, what is only potentially affected, what is shared, and which edges CORVEN cannot see or enforce.
Blast radiusContain
A minimum-cost cut over enforceable edges that disconnects the compromised agent from protected resources while keeping declared healthy workflows intact. A second person approves the exact plan.
See a planVerify
Fresh probes through the gateway: the compromised path must now be denied, healthy paths must still be allowed. Partial results stay partial.
See verificationProve
The whole response is sealed as SHA-256-hashed, chained evidence, stored on Arkiv, anchored on Ethereum Sepolia and open to an independent Chainlink CRE re-check.
See evidenceThe architecture
A control plane, an enforcement point, and a trust layer.
The control plane holds the registry, the dependency graph and the incident workflow. The gateway is where decisions are enforced: CORVEN can only block calls that pass through it, and says so for every edge that does not. Evidence leaves CORVEN through the trust layer, so it can be checked without trusting CORVEN's database.
- 01DiscoveryContinuous mapping of agents, resources & delegated authority.
- 02Graph EngineDeterministic blast-radius traversal & cycle detection.
- 03ContainmentCalculates the minimum safe cut boundary.
- 04GatewayAuthorization enforcement proxy with dynamic access revocation.
- 05ApprovalsCryptographically bound human authorization before enforcement.
- 06EvidenceTamper-evident incident proofs sealed on Arkiv & Ethereum.
Integrations
Identity, storage and an independent check.
◎ENS · Ethereum Sepolia
Human-readable identity.
Agents and operators can carry ENS names. CORVEN resolves names to addresses and addresses back to primary names, so evidence and timelines read as names, not hex. A name does not make an agent trustworthy and never authorises anything.
Live: forward and reverse lookup in the console (Agents → ENS lookup).
Try a lookup▣Arkiv · Rabdi testnet
Evidence that lives outside CORVEN.
Every sealed evidence record is written to Arkiv as an entity with its hash and references, then read back and compared field by field. It is shown as stored only after the read-back matches.
Live: each demo run publishes its evidence to Rabdi (chain 7794095) and shows the entity and transaction.
Arkiv explorer⬡Chainlink CRE · Ethereum
An independent re-check.
The evidence hash is anchored in the EvidenceAnchor contract on Sepolia. A Chainlink CRE workflow re-hashes the evidence and checks it against that anchor; CORVEN accepts only results signed by the Chainlink network and verified against its registry.
Anchor: 0xef716752f53ad089789bc8b57f68fbf00357ee1b (Sepolia, verified source).
CRE: workflow built and tested; deployment to the Chainlink network is pending approval, and the console shows it as not configured until then.
What these prove: a matching hash proves an evidence record has not changed since it was sealed. It does not prove the recorded events were true, and an anchor is counted only after it is confirmed on chain.
Under the hood
The engines behind dependency-aware response.
∑Graph engine
Deterministic blast radius.
Reachability, path and cycle analysis over the registered dependency graph, with a reason for every affected node and an explicit list of what it could not see.
Engine spec◇Containment
Minimum safe cut.
The smallest set of enforceable revocations that separates compromised agents from protected resources, checked against the healthy workflows you declare.
How it contains⌘Verification
Checked, not assumed.
After execution CORVEN probes again. An incident is marked contained only when the denied path is denied and the healthy path still works.
Verification- TypeScript
- Fastify
- PostgreSQL
- Next.js
- Foundry
- ENS
- Arkiv
- Chainlink CRE
No account needed
See what happens when an agent gets compromised.
The demo creates an isolated sandbox of five connected agents and walks one incident from analysis to sealed, anchored evidence on the real CORVEN engines.