i̲s iAgentSafe
Part of the iJarvis Agent Stack

Stop your agent before it errs.
Prompt injection defense + counterparty risk checks + kill switch.

Pre-action verification middleware for autonomous agents. Detects prompt injection attempts, validates counterparty risk via iAgentFi, monitors behavior anomalies against baseline, and provides a kill switch when something goes wrong.

iagentsafe.check(action, context) -> allow | deny | review

iAgentSafe will provide four primitives

Prompt injection detection

Pattern library plus LLM-classifier on incoming instructions. Detect jailbreaks, context poisoning, indirect injection via tool outputs.

Counterparty risk policies

Integrate iAgentFi ratings. Refuse actions with counterparties below policy threshold. Configurable per action type.

Behavior baseline

Learn normal agent behavior via iAgentLog. Alert on anomalies: unexpected tool calls, unusual spending, off-goal actions.

Remote kill switch

One API call halts all actions for an agent. Emergency stop for supervised autonomous systems.

The gap iAgentSafe fills

Libraries exist; integrated platform doesn't

Guardrails AI, NVIDIA NeMo, Lakera are all library-level. iAgentSafe will ship as the integrated service with trust signals + behavior baselines + real-time enforcement (Cluster 2 roadmap 2027).

Most attacks happen at runtime

Prompt injection, indirect injection, context manipulation: these happen during agent operation, not at deploy time. Runtime guardrails are the only defense.

Open-core model

Core SDK open-source under MIT. Hosted service with integrations, managed policies, and incident response is paid.

Planned endpoints in Roadmap 2027 (Cluster 2)

Preview of the planned API surface. OpenAPI 3.1 specification at /.well-known/openapi.yaml. Endpoints at api.iagentsafe.com will serve requests at roadmap 2027 (cluster 2); agent-consumable JSON by design.

POST /v1/check Pre-action verification: allow, deny, or review
POST /v1/policy Configure runtime policies
POST /v1/killswitch/{agent_id} Emergency halt for a specific agent
GET /v1/incidents/{agent_id} Security incident history
Status: Cluster 2 roadmap 2027. Stack-dependent product; launches after Cluster 1 customers pull for runtime enforcement.

iAgentSafe is one layer

Sixteen products. One stack. One entity. Trust, discovery, observability, payments, safety, simulation, composition, memory, identity, legal, markets, and owned compute underneath. Each layer reinforces the others. Use one or use them all.