[init] session=int-7f3a   environment=sealed   network=none

~/agent_security/failure_lab

Test consequential
behavior without consequential
exposure.

Sandboxed attack simulation for agentic systems. Exercise models, tools, identities, protocols, and data inside restorable worlds—then trace the recorded path from adversarial input to consequential effect.

./ inspect the architecture
interceptor://int-7f3a nearest: MCP_SERVICE
  • 01 DNS
  • 02 HTTP
  • 03 MCP
  • 04 local_tool
  • 05 attacker_sink
  • 06 filesystem
  • 07 app_interface
  • 08 postgres
INTRUSIVE.AI, Inc. sandboxed attack simulation evidence + restore session / active

./ overview.md

## attack_surface

scope: agentic_systems

The attack surface is
larger than the prompt.

Agentic systems do more than generate output. They select tools, assume identities, retrieve data, call services, change state, and create external effects.

The paths worth testing are often the paths least safe to exercise against production. INTRUSIVE.AI replaces the surrounding world with observable, restorable simulations so testers can explore those paths without inheriting their blast radius.

  1. 01adversarial inputprompt / document / event
  2. 02modelinterpretation
  3. 03plangoal + decisions
  4. 04tool callcapability use
  5. 05state changesystem mutation
  6. 06external effectconsequence

./ attack-surface.md

## architecture

architecture: 2_domains + 1_loop

Architecture for controlled attack simulation.

Separate the offensive system from the target system, bound the authority of each, and let observed behavior determine what happens next.

Intrusive Threat Intelligence plans scenarios and assesses results. Operator Sandbox governs the adaptive Attack Harness and mediates its access to the target application and synthetic world in Interceptor Sandbox.
domain::operator attack harness
intrusive threat intelligence
domain::interceptor agent + synthetic world

operator_sandbox.md

state: contained

Contain the tester.

Operator Sandbox

Runs offensive AI testing inside a disposable execution boundary. The host mediates model and target access through explicitly configured operations, credentials, and limits.

accepts
attack harness + authorization policy
governs
execution + capability access
emits
attempted actions + operator trace

intrusive_threat_intelligence.md

state: adapting

Expand the search.

Intrusive Threat Intelligence

Builds threat intelligence informed scenarios based on the target agent's environment and capabilities and analyzes the combined sandbox results of the simulated attack campaign.

accepts
target agent capabilities
governs
threat model + simulation scenarios
emits
attack simulation performance analysis

interceptor_sandbox.md

state: observable

Contain the target.

Interceptor Sandbox

Places a real agent application inside a synthetic operating world. Services, identities, data, tools, and attacker infrastructure remain inspectable, with declared state restorable. Supported model calls pass through a host-side connector.

accepts
agent build + Environment Blueprint
governs
state + services + external effects
emits
recorded observations + logical snapshots

./ system.md

## mechanics

scope: execution

How the sandboxes work.

Choose a sandbox to see what runs inside it, how it communicates, and what gets recorded.

isolated tester / host-mediated access

the testing side

Operator Sandbox

Runs the AI-powered Attack Harness inside a network-isolated container. The trusted host controls every permitted connection to a model or target.

  1. Define the investigation.

    Supply objectives and optional scenarios. Operator checks the inputs against target capabilities, configured permissions, and limits.

  2. Start the isolated tester.

    Attack Harness plans experiments and adapts to feedback using a tool catalog fixed before the campaign starts.

  3. Broker the permitted actions.

    Operator makes model and target requests on the harness’s behalf. Credentials and target destinations stay under host control.

  4. Retain the results.

    Review the accepted inputs, attempted actions, observations, and assessment together, including missing evidence and untested work.

boundary // Attack Harness cannot change its tool configuration or choose a different target destination.

./ read the Operator guide ↗
contained target / controlled dependencies

the application side

Interceptor Sandbox

Runs your agent in a container with the services, files, identities, and data declared for its environment. Observe how it behaves when those inputs change.

  1. Describe the agent’s world.

    An environment blueprint defines how to invoke the agent and configures its dependencies and starting data.

  2. Run against controlled dependencies.

    Provide supported HTTP, MCP, PostgreSQL/pgvector, and file surfaces. Real model calls and declared HTTPS seams use configured host-side connections.

  3. Place payloads at the boundary.

    Use the injection API to alter supported model, tool, service, or database responses, or declared mutable environment state.

  4. Observe and compare behavior.

    Inspect recorded interactions and state changes. Where snapshot prerequisites are met, restore declared logical state to compare experiments.

boundary // Declared connections do not grant general network access. Snapshots do not capture arbitrary process memory or remote service state.

./ read the Interceptor guide ↗

./ mechanics.md

## use_cases

scope: practitioner_workflows

Security evaluation for agentic systems.

Select a use case to explore what to evaluate and the question it helps answer.

./ use-cases.md

## scenarios

coverage: system_level

Scenarios,
not prompt lists.

A successful jailbreak is not always a consequential failure. A system fails when adversarial influence crosses a boundary and changes what an agent can observe, decide, access, or affect.

RAG / INPUT

Indirect prompt injection

Retrieved content rewrites the agent’s goal and drives an unauthorized action.

MCP / AUTHORITY

Capability escalation

A poisoned or overpowered tool expands what the agent is permitted to do.

IDENTITY / ACCESS

Credential misuse

The agent crosses a user, service, or role boundary while completing a workflow.

PROTOCOL / EGRESS

Data exfiltration

Sensitive state reaches an attacker-controlled HTTP, DNS, or tool endpoint.

TOOL / EXECUTION

Unsafe local action

A legitimate local capability is invoked outside its intended scope.

AGENT / GOAL

Multi-turn goal hijacking

Escalating pressure moves the agent from its authorized objective to the attacker’s.

MULTI-AGENT / TRUST

Role contamination

Instructions or authority bleed across agent roles and coordination boundaries.

RUNTIME / BUDGET

Resource exhaustion

Loops, retries, or tool chains consume resources beyond defined stopping conditions.

./ scenarios.md

## open_source

status: open_source

operator + attack_harness + interceptor / AGPL-3.0-only

Built to be inspected.

Operator Sandbox, Attack Harness, and Interceptor Sandbox are open source software licensed under the GNU Affero General Public License version 3 only (AGPL-3.0-only). Their execution boundaries, synthetic services, and isolation mechanisms remain open to inspection and practitioner extension.

release_authority://public Canonical release destinations, artifacts, and validation records. inspect releases
operator
open source / AGPL-3.0-only
attack harness
open source / AGPL-3.0-only
interceptor
open source / AGPL-3.0-only
participation
practitioner-driven

./ open-source.md

## mission

independence requires evidence //

To be the world's trusted source of safe-to-use intrusive AI capabilities.

Exhaustively test everything we build.

Radically transparent about the results.

./ mission.md

## start

audience: practitioners

Bring us an adversarial scenario that can't be left to chance.

We are working with AI security testers, red teams, and application builders on high-consequence evaluation problems.

./ start a technical conversation hello@intrusive.ai ↗

./ start.md