🤖
Agents with tools
The agent calls APIs, writes to the database, sends email or touches infrastructure. The question isn't whether it goes wrong — it's what it can do when it does.
Zive AI Security · a Zive IT division
Security for AI agents, apps and the APIs behind them
Engineering-depth reviews · reproducible findings · practical fixes
16+ years building mobile, backend, cloud and AI systems — now applied to assessing what those systems actually allow. Tool permissions, server-side authorization, tenant isolation, prompt injection with real consequence and data leakage. With written authorization, a fixed timeline and a retest.
16+
years of software engineering
3–10
business days per assessment
100%
of findings reviewed by hand
14d
for the retest, included
Who it's for
When an AI agent is already in production
Teams that shipped agents, copilots or AI automations with access to internal data and real tools — and need to know what that access allows before someone else finds out for them.
🧩
RAG and internal data
Semantic search over company documents, multiple customers in the same index, third-party content entering the model's context.
📱
Mobile app + API
A published Android or iOS app and the API behind it. Authentication, object-level authorization, session, local storage and deep links.
🏢
Multi-tenant SaaS
Many customers on the same base. Tenant isolation is the kind of failure that rarely shows up in automated testing.
🚀
Before a launch
A new AI feature going live, investor technical due diligence, or an enterprise customer asking for assessment evidence.
🤝
White-label partnership
Cloud consultancies, software studios and AppSec boutiques that need a technical partner for specific assessments.
Assessments
Three formats, fixed scope
Each assessment has scope, timeline and deliverables defined in writing before it starts. No open-ended projects.
Most requested
AI Agent Security Sprint
Up to 1 agent and up to 3 critical integrations or tools.
- Threat model of the agent flow
- Controlled adversarial evaluations
- Permission and authorization review
- Technical and executive report
- One targeted retest included
US$ 4.000–7.500
5–8 business days
Mobile + backend
API + Mobile Security Review
1 Android or iOS app and up to 10 critical API endpoints.
- Authentication and object-level authorization
- Local storage and secrets in the app
- Deep links, IPC and exported components
- Session and backend controls
- Limited retest included
US$ 5.000–9.000
7–10 business days
No intrusive testing
AI Security Architecture Review
White-box architecture review, no intrusive exploitation.
- Access control and identities
- RAG, data and trust boundaries
- Secrets and infrastructure
- Logging and audit trails
- Prioritized fix backlog
US$ 2.500–5.000
3–5 business days
The ranges above are indicative: the final figure depends on scope, access, environment and liability, and is fixed in the SOW before work starts. An architecture review and a full penetration test are different services and are contracted separately. For monthly reviews, change retests and configuration review there is a retainer from US$ 1,500/month.
Standard scope
What the AI Agent Security Sprint covers
5 to 8 day sprint
From user input to side effect
The assessment follows the whole path an instruction travels — user input → retrieval/RAG → model → tool selection → authorization → side effect — and looks for where that path trusts something it shouldn't.
- Inventory of agents, models, tools, MCP servers, identities and reachable data
- Controlled direct and indirect prompt-injection tests, with concrete impact criteria
- Server-side authorization checks, least privilege and approval for sensitive actions
- Isolation between users and between tenants
- Context manipulation, data leakage and credentials inside tools
- Logging and audit trails: can you reconstruct what the agent did?
- Severity by probability, required privilege, exposure and real impact — not by scanner score
- A 60-minute closing meeting and one targeted retest within 14 days
Deliverables
What you keep at the end
A scanner produces a hypothesis; a serious finding needs impact and evidence validated by a human. Everything in the report was reproduced by hand.
2-page executive summary
Risk in business language, for whoever decides budget and timeline.
Technical report
Per finding: ID, impact, preconditions, evidence, mitigated reproduction and recommendation.
Risk matrix
Severity by probability, privilege, exposure and impact in your context.
Prioritized fix backlog
In execution order, with control examples — ready to become tickets on your board.
CTO summary
What to change in architecture and controls, not just what to patch now.
Retest record
A record that the fix was verified — the document your customer or auditor asks for.
Process
How it works, from first contact to retest
Discovery call (25 min)
What's in production, which tools the agent can reach and what the real concern is. No commitment.
Technical questionnaire
Asset inventory and risk classification, to size scope and timeline with numbers rather than guesswork.
NDA, SOW and written authorization
In-scope assets, permitted and prohibited methods, windows, request ceiling, emergency contacts and kill switch.
Kickoff and test accounts
Agreed environment, synthetic accounts and test data. Production and real data only with express permission.
Controlled assessment
Testing within the authorized scope, with safe reproducible evidence. No intrusive action without human approval.
Report, fix and retest
Documents delivered, a 60-minute fix meeting and one targeted retest within 14 days.
Payment terms: 50% on SOW acceptance and 50% on delivery, with written milestones and acceptance criteria.
Limits and authorization
What this service is — and isn't
Security sold on a promise that can't be kept is a risk, not a control. So it's written here, before any proposal.
What I promise
- Identify exploitable risks and control failures in a verifiable way
- Safe, reproducible evidence for every finding, reviewed by a human
- Help you fix it: recommendations with control examples and priority order
- Work only within the signed scope and authorization
- Confidentiality, per-client segregation and evidence disposal on the agreed schedule
What I don't promise
- Total absence of vulnerabilities — no assessment delivers that
- Automatic formal compliance, certification or regulatory accreditation
- Guaranteed discovery of flaws in a given scope
- Full network penetration testing, OT/ICS or ransomware simulation
- 24/7 SOC monitoring — without a dedicated team and SLA, that isn't for sale
Operating rule: deny by default. Nothing is tested without express written authorization. No DoS, no persistence, no phishing or social engineering, no extraction of real user data, no lateral movement and no testing of third parties. AI is used to speed up triage and writing, never to decide to expand scope or run an intrusive action — that stays a human decision.
Method
Public references, not a secret methodology
The assessment is mapped to public checklists, so you can verify what was covered — and what wasn't.
AI and agents
OWASP Agentic Top 10OWASP LLM Top 10PromptfooPyRITMCP
API and web
OWASP WSTGBOLA / IDORMulti-tenancySSRFBurp SuiteOWASP ZAP
Mobile
OWASP MASVSMASTGMobSFKeychain / KeystoreDeep links
Code and dependencies
SemgrepCodeQLGitleaksTrivyCheckov
Cloud and identity
AWS IAMLeast privilegeContainersSupply chainLogging
Threat modeling
STRIDEAbuse casesCVSSBusiness context
Get in touch
Let's look at your scope
Tell me what's in production and what the concern is. The first call is 25 minutes and exists to decide whether this makes sense — I reply personally.
WhatsApp
+55 (41) 99616-2887
E-mail
arildo.junior@ziveit.com
LinkedIn
linkedin.com/in/arildo
Location
Curitiba, Brazil · remote engagements