AI security and penetration testing
Find out how your AI assistants and agents can be misused before an attacker does, and fix it before launch. We test web apps, networks, cloud and physical sites too.
Security that follows the work, not only the data.
Access to data doesn't authorize every AI action. We secure the whole workflow, then prove the controls hold with AI penetration testing.
Where an AI workflow can go wrong
Select a stage to see the risk and our control.
Risk
Instructions hidden in emails, documents or web pages hijack the assistant (prompt injection).
Control
Instruction hierarchy, input provenance checks and adversarial testing against realistic attacks.
Risk
Retrieval pulls sensitive records into the model's context beyond what the task needs.
Control
Classification-aware retrieval, sensitivity labels and DLP applied to AI interactions.
Risk
The model is talked out of its restrictions or plans an unsafe course of action.
Control
Policy guardrails, evaluation suites and red-team findings tied to release decisions.
Risk
Over-privileged tools let an agent change production systems or send data out.
Control
Least-privilege tool scopes and human approval for irreversible actions.
Risk
Sub-agents inherit authority that nobody reviewed or owns.
Control
Scoped delegation, named owners and traceable handoffs between agents and people.
Risk
Automated actions that nobody can explain, audit or roll back.
Control
Audit trails, continuity and recovery plans, and evidence that supports sign-off.
Standards we map our work to
- OWASP Top 10 for LLM Applications
- MITRE ATLAS
- NIST AI RMF
- ISO/IEC 42001
- ISO/IEC 27001
- EU AI Act
- Israel Privacy Protection Law, Amendment 13
What you receive
Every assessment ends with findings your engineers can fix and a summary your board can read.
Every engagement includes
- Documented findings with clear fixes
- A one-page board summary
- A retest of fixed findings
Support bot can be talked into issuing refunds
- Mapped to
- OWASP LLM06 Excessive Agency, MITRE ATLAS AML.T0051 LLM Prompt Injection
- Evidence
- A crafted customer message made the agent call the refund tool without approval.
- Fix
- Require human approval for refunds and limit the tool to read-only order lookups.
- Retest
- Passed after the fix
What we cover
Secure architecture and governance for the AI you build or buy, and penetration testing across AI, applications, networks and physical sites.
AI security, governance and privacy
Controls that protect sensitive data and keep automated actions accountable, without blocking useful work.
Includes
- AI threat modeling and secure architecture
- Secure RAG, data leakage and prompt-injection risk
- AI governance, policies and risk classification
- Privacy, compliance and third-party AI risk
Penetration testing and red teaming
We attack your systems the way real adversaries would, from AI assistants to the front door, and hand you documented findings with clear fixes.
Includes
- AI and LLM application pentesting: prompt injection, jailbreaks, agent and tool abuse
- Web, API, mobile and cloud penetration testing
- Network and infrastructure testing, internal and external
- Physical intrusion and social engineering, including phishing
How to start
Three fixed-scope ways to begin. Each ends with a clear report and a plan for what comes next.
AI security assessment
Typical duration: 2–3 weeks
One AI assistant, agent or workflow reviewed end to end: threat model, adversarial tests and data-exposure checks.
Ask about thisYou receive
Findings with evidence, a prioritized fix plan and a board summary.
Microsoft 365 Copilot readiness
Typical duration: 3 weeks
Find what Copilot could expose before your employees stumble on it, then fix permissions, labels and DLP.
Ask about thisYou receive
An oversharing map, a label and DLP plan, and rollout guardrails.
Penetration test
Typical duration: 1–4 weeks, scoped
AI applications, web and API, network and cloud, or physical intrusion and social engineering, scoped to what matters most.
Ask about thisYou receive
A report with evidence and clear fixes, and a retest once you have fixed them.
Not sure where to start? Begin with a 30-minute discovery call.
Bring us one AI challenge that matters.
Most engagements start with a discovery call or a focused review of one workflow. Tell us what you are working on.
Or reach us directly
Derech Menachem Begin, Tel Aviv-Yafo, IsraelWorking with clients in Israel and worldwide.