Independent security engineeringEurope based · US LLC

Security at the
point of action.

An AI agent can read, decide and act.
We help you define where its authority ends, test the boundary and build the controls.

Discuss your system ↗
AI applications, agent workflows, APIs and cloud infrastructure.
Explore an execution boundary

Allowed. The request stays within its declared permissions.

Interactive illustration. No live agent or tool execution.
For SaaS teams building with agents

AI agent security assessment.

Test whether your agent can access data, invoke tools or take actions beyond its intended permissions.

For teams connecting agents to customer data and business APIs, preparing for launch or reviewing an existing workflow.

What you receive

Reproducible findings, supporting evidence and practical fixes your engineers can work from.

We agree the workflow, environment and test boundaries before starting. Timing and fees follow the scope.

01 / Assess the exposure02 / Implement the controls03 / Verify the result
01 / Engagements

A defined problem.
A practical deliverable.

Assessment / AI

AI application & agent security

For teams connecting models to business data, APIs and tools. Review trust boundaries and test prompt injection, data access, tool permissions and approval behaviour within an agreed scope.

You receiveReproducible findings, a prioritised remediation plan, an engineering walkthrough and an agreed retest.

Explore the engagement

Example assessment scope

One AI application, its retrieval sources and connected tools. Agree the test environment, accounts and permitted actions before testing.

Illustrative starting point

A support agent that retrieves customer records and can update tickets.

Find your starting point ↗
Implementation / AI

Secure AI adoption

Take one useful workflow from idea to a bounded pilot. Define the data access, tool permissions, human approvals and evaluation criteria, then implement and test the integration.

You receiveA scoped pilot, security and reliability evaluations, documented limitations and an operating handover.

Explore the engagement

Example pilot scope

One workflow with explicit data sources, tool permissions, approval points and success criteria. Start in a controlled environment before expanding access.

Illustrative starting point

An internal assistant that prepares a draft report and waits for approval before publishing.

Find your starting point ↗
Engineering / AppSec

API & cloud security

Review API authorization, Azure infrastructure and delivery controls. Work with your engineers to investigate findings, implement fixes and add checks to the release process.

You receiveValidated findings, practical control changes and verification evidence. Project work or an agreed ongoing allocation.

Explore the engagement

Example engineering scope

One API or cloud workload. Review authorization, identities, deployment configuration and the controls protecting its data and operations.

Illustrative starting point

A multi-tenant API where each customer must only access its own records.

Find your starting point ↗
02 / A boundary, made concrete

The model asks.
The control decides.

A convincing instruction is not permission. Choose a request to see how this simplified illustration checks a tool call against a declared file-access grant.

Illustrative interaction only. No agent runs and no files are accessed.

CALL RECORD / ILLUSTRATION
REQUESTED ACTION
files.read → notes/project.md
DECLARED GRANT
files.read → notes/**
CONTROL CHECK
The resource is inside the permitted scope.
ALLOWED / proceed to tool
03 / Independent engineering

Inspect the thinking.
Read the code.

Python / Alpha

Legion

An execution-control runtime for agent tool calls. Scoped grants, call-bound approvals, narrowing delegation and recovery designed around uncertain effects.

Explore the repository ↗
Go / Agent security

NIA

Credential-bound agent identity, authorization, runtime monitoring and revocation. REST and MCP requests pass through a shared enforcement path.

Explore the repository ↗
Python + C# / Early development

Witness

Evidence-based triage of scanner findings in supported C#/.NET vulnerability classes. Deterministic checks preserve uncertainty when the evidence is insufficient.

Explore the repository ↗

Independent open-source projects. They demonstrate engineering work and are not presented as client deployments.

04 / Find a starting point

What are you
building?

Suggested first conversation

AI application & agent assessment

Start with the data and tools your system can reach. Review permissions, prompt injection exposure and actions that require human approval.

A defined test scope, reproducible findings and prioritised remediation.

Discuss this scope ↗This is a starting point. Final scope follows a conversation about your system.
How an engagement works

Defined scope.
Evidence you can use.

01 / Scope

Agree what we will test.

Map the workflow, data, identities and tools. Agree the environment, test accounts, permitted actions, exclusions and success criteria before testing starts.

02 / Assess

Trace authority to the action.

Review architecture and permissions, then test the agreed attack paths. Check the resulting effects and logs as well as the model’s response.

03 / Deliver

Make each finding actionable.

Receive an executive summary and technical findings with affected components, reproduction steps, evidence, impact, recommended fixes and coverage limitations.

04 / Verify

Check the fix against the failure.

Walk through findings with your engineers. Agree remediation support and retesting separately, with clear checks for each control change.

Access depends on scope: typically architecture documentation, relevant configuration or code, a test environment, dedicated accounts and logs. Use synthetic data where possible. Timing and fees are agreed after scoping.

See the work

From a claim
to a test.

Illustrative deliverable

A tool can cross a tenant boundary.

Read a sample finding, its supporting evidence, the proposed fix and the checks needed to verify it.

Read the sample assessment ↗
Technical walkthrough

Can your agent read another tenant’s data?

A small reproducible Python example showing why a prompt instruction cannot replace authorization at the tool boundary.

Follow the walkthrough ↗

Explore permissions, revocation and approval in the 3D lab ↗

05 / Start with the system

What are you
giving AI access to?

Start with a short description. We will discuss whether an assessment or implementation engagement fits.

  • What your system does and its deployment stage.
  • The data, APIs and tools it can access.
  • Your main concern and any target date.

No credentials or sensitive customer data are needed for the first enquiry.

Founded by Bogdan Ticu, a security engineer working across AI, application and cloud security.

Remote delivery from Europe.
B2B contracting through Bitwise Defence LLC, New Mexico.

A few details to start.

This form prepares an email. Review and send it in your email app. Nothing is submitted or stored by this website. If you use webmail, choose “Copy enquiry” and paste it into a message to contact@bitwisedefence.com.

Please leave out passwords, API keys and sensitive customer information. Enquiries sent by email will be used to discuss your request.

← Back to the website

Sample assessment / Fictional system

One tool.
Two tenants.
A missing boundary.

What an actionable agent security finding can look like, from the requested action to a verifiable fix.

Demonstration, not a client result.

This fictional support agent and synthetic evidence illustrate a deliverable. No production system was assessed. Severity would depend on the actual data, exposure and compensating controls.

Scope and intended behavior

A support assistant retrieves invoices for the authenticated customer. The test covers one read-only tool, two synthetic tenants and the authorization boundary between the agent and the invoice service. Billing changes, infrastructure and other tools are excluded.

Finding BD-DEMO-01

Cross-tenant invoice access through a model-controlled argument. The tool accepts a tenant identifier supplied in the proposed call. The backend uses that value without binding it to the authenticated customer. A changed argument can return another tenant’s invoice.

Expected: the authenticated tenant A can read only tenant A invoices.

Requested: get_invoice(tenant="B", invoice="INV-002") under tenant A’s session.

Vulnerable fixture result: tenant B’s synthetic invoice is returned.

Impact: unauthorized disclosure if the same boundary is missing in a real deployment.

Reproduction

  1. Create isolated synthetic tenants A and B, each with an invoice.
  2. Authenticate as tenant A and confirm that its own invoice is accessible.
  3. Propose a tool call for tenant B’s invoice using the same authenticated session.
  4. Capture the tool result and backend access log. Establish whether the foreign record was actually returned.

A deterministic call is enough to establish this authorization defect. Prompt injection could influence the proposed arguments, but its success is a separate test. A refusal in model text would not demonstrate backend enforcement.

Recommended fix

Derive the tenant from trusted authentication context. Reject conflicting arguments and scope the database lookup to that tenant. Enforce the same rule at the backend even if a gateway also checks it. Record the authenticated principal, resource and decision without logging invoice contents.

Retest acceptance criteria

Coverage and limits

This finding addresses tenant isolation for one tool. It does not establish protection against every prompt injection, approval replay, credential compromise or data path. A full assessment records tested paths, excluded components and remaining uncertainty.

Run the small example ↗ Discuss your system ↗
← Back to the website

Technical note / Local synthetic example

Can your agent read another tenant’s data?

Treat the proposed tool call as input. Bind access to the authenticated principal at the point where data is returned.

Imagine a support agent told to serve only its current customer. Its invoice tool still accepts a tenant argument. If the backend trusts that argument, the instruction offers no enforced tenant isolation.

Reproduce the boundary failure

Save this as tenant_boundary_demo.py and run python3 tenant_boundary_demo.py. It uses only the Python standard library and synthetic data. No network calls or model are involved.

Expected output

Vulnerable: synthetic invoice B
Protected: denied
Own tenant: synthetic invoice A

The first call returns a foreign invoice. The protected version checks the requested tenant against trusted session context before looking up the record. Its assertion also verifies that the legitimate read continues to work.

What this does and does not prove

This is an authorization test, not a prompt injection benchmark. It demonstrates the effect of a proposed out-of-scope tool call. Whether an attacker can persuade a particular model to propose that call needs a separate behavioral test.

Apply it to a real system

A tenant check is only one part of the design. Production systems also need proper authentication, resource permissions, input validation, audit handling and controls appropriate to each action.

See the sample finding ↗ Discuss your workflow ↗
← Back to the website