Web & API pentest
Manual testing of web applications and APIs, beyond what scanners report: access control, business logic and the way components chain together.
- OWASP Top 10 / API Top 10
- Authn, authz & IDOR
- Business-logic abuse
$ whoami
Pentest/Red team/Offensive security/Vuln research
I'm Giulliano Amaral, a penetration tester and security researcher. I find the paths an attacker would take through your applications, networks and AI systems, then document them clearly enough for your team to close them.
./why --pentest
Because someone is going to test your company's security. The only question is whether it's someone you hired, or an attacker.
Most breaches don't need sophisticated tradecraft. They start with a known weakness nobody tested, in an exposed system or a supplier. The latest industry data shows how big the problem is:
is the average cost of a data breach, a record high and 12% more than the year before.
of breaches started with the exploitation of a vulnerability, now the most common way attackers get in.
is the median time to fully fix a critical, actively exploited vulnerability. Only 26% of them were fully remediated.
of breaches involved ransomware, and another 48% involved a third party such as a supplier or partner.
If exploiting vulnerabilities is the top way in, the most direct defense is to look for them the way an attacker would, with manual testing of your applications, APIs and network.
Teams fix slowly because they have too many findings. The report ranks each one by real impact and exploitability, so effort goes to what actually opens the door.
An independent report answers customers' security questionnaires and supports audits such as ISO 27001 and SOC 2, and shows due diligence under data-protection laws like GDPR and Brazil's LGPD.
cat services.txt
Every engagement is manual, scoped in writing, and ends with a report your developers can act on: reproducible steps, impact and fixes.
Manual testing of web applications and APIs, beyond what scanners report: access control, business logic and the way components chain together.
Assume-breach assessments that start from a foothold and map the realistic path to domain compromise.
Adversarial testing of LLM apps and agents: prompt injection, tool abuse and trust boundaries between agents.
ls -t writeups/
End-to-end attack chains from HackTheBox machines and AI red-team labs. Active machines stay sealed until they retire.
A calculator agent that runs your input through eval() — bypassing an input denylist and a numeric-only output filter by encoding a file as one big integer.
A connectivity tool drops host straight into a shell and only says Online/Offline — so read the flag out-of-band through the server's own access logs.
Active HTB Season 12 box — rooted. An LLM text-to-SQL chatbot leaks its own operator override, turning read-only queries into arbitrary SQL; impersonate a sysadmin login, enable xp_cmdshell for a shell as the SQL service account, then decrypt a SQL Agent proxy credential (Service Master Key, with a LEN→DATALENGTH fix) to reach a Backup Operators account and read root via SeBackupPrivilege. The full chain is sealed (password-protected); reach out if you need access.
Active HTB Season 12 box — rooted. Go/Echo web app: SSTI into the ticket PDF, a forged signed QR, an internal automation service and a container/host boundary. The full chain is sealed (password-protected); reach out if you need access.
Active HTB Season 12 box — rooted. The full chain is sealed (password-protected); reach out if you need access.
ls articles/
Advanced search operators turn Google into the cheapest recon tool you own. The operators, a copy-paste OSINT cheatsheet, how to defend against it — and real incidents where a search query was the first step of the attack.
The API Red Team exam, written up defender-first: a JWKS endpoint that leaked the HS256 secret, token forging into IDOR and role escalation, client-trusted pricing, a Lambda /debug that dumped the cloud, an SSRF-into-S3 object read, and a UNION SQLi for the rest. Sanitised — live card number, gift-card code and secret stay out.
The Certified AI Red Team Analyst exam, written up defender-first: an MCP-backed docs copilot, a RAG knowledge base I could write to, an agent-only SSRF tool, a leaked ssh-tunnel account, and an internal agent fleet with one over-trusting document agent. Sanitised — the live flag and exam creds stay out.
The jailbreak pipeline — recon, craft, deliver, bypass — the technique families behind it, and the defence-in-depth that actually breaks the chain. Defender-first, sanitised PoCs.
ls ~/certs
Hands-on, lab-based exams covering pentest, red team, Active Directory, web, API and AI security.
./contact --scope yours
Tell me about the target, the environment and your timeline. I'll come back with questions and a proposed approach. Testing only starts with written authorization.