Oktsec Assessment · AI Security
Pentesting for AI agents, web applications, APIs and infrastructure, with evidence and a clear remediation plan.
Based in Argentina, serving Latin America. Before a launch, an audit or a customer security review.
250+ security findings reported
Microsoft
Cloudflare
Mercury
gbrain / gstackxAI
The experience behind Oktsec
Cybersecurity entrepreneur focused on AI agent security. Over 20 years in technology, including 12 as CTO. Ambassador for the Agentic AI Foundation (AAIF), part of the Linux Foundation.
Founder’s research · 2026
Two contributions by Gustavo Aragón to Mercury and Stripe tools, with public reports, code and tests.
Mercurymercury-cliDebug mode redacted sensitive request data but left response data visible. Cookies could end up in automation logs or shared reports.
Redact sensitive response headers as well, with tests to verify the fix.
link-cliOn a shared filesystem, a link prepared by another user could redirect the payment credentials file to a location that user could already read.
Create the file exclusively with restricted permissions and avoid writing through symbolic links. This protection applies to systems such as Linux and macOS.
Contributions by @garagon merged into the official repositories. This is public research by the founder, not work commissioned by these companies.
He also reported findings and proposed fixes in gbrain ↗ and gstack ↗, open source projects by Garry Tan, president and CEO of Y Combinator.
What we test
We assess your applications, agents and infrastructure to identify risks and define how to fix them.
| Area | What we test | What we look for |
|---|---|---|
| AI agent pentesting | AI red teaming for LLM applications, MCP tools and instructions received by agents. | Data exposure, task hijacking and actions beyond authorized permissions. |
| Web application and API pentesting | Authentication, sessions, access controls and business logic. | Access to other customers’ information and operations without the required permissions. |
| Repositories and architecture | Code, dependencies, integrations and system design. | Vulnerabilities that can be chained to compromise the system. |
| Infrastructure, credentials and permissions | Cloud configuration, secrets management and access permissions. | Exposed credentials, accessible resources and excessive permissions. |
After the first pentest
An initial pentest followed by assessment cycles as your agents, applications and infrastructure evolve.
Scope a recurring assessment ↓The proposal defines the frequency, systems covered and changes that trigger a new assessment.
Before requesting a proposal
Choose your systems, size and goal. Then send us your scope to receive a proposal.
Prefer to talk directly?
Tell us what you need
What you receive
To address risks and report to leadership, customers and auditors.
Disclosures from August and September 2026
Metabase confirmed an attack that could allow administrator access, theft of connected database credentials and data export. It released patched versions.
Wiz demonstrated that a GitHub Actions flaw could expose a credential and allow access to internal information. Snowflake fixed the workflow and revoked the credential; the disclosure describes a research test.
Manifold documented flaws in coding agents when opening folders containing malicious Git configuration, such as folders received in a ZIP. Some executed code before user approval. Fix status varies by agent.
Third-party cases with original sources. Dates refer to public disclosure. These are not Oktsec clients or findings.
How to get started
Tell us which system you want to assess, its integrations and your assessment goals.
You receive the scope, exclusions, schedule and price. We begin with your authorization.
We run the tests and review findings with your team to prioritize fixes.
We work remotely, on site at your offices or through a hybrid arrangement. We agree location, access and evidence handling around your company policies; the proposal confirms availability and logistics.
Before you start
Yes. We can arrange remote pentesting and assessments, on-site testing at your offices or a hybrid approach. If your policy restricts remote access or information leaving your environment, we agree how to work within it. The proposal confirms location, availability, access and logistics. For on-site or hybrid engagements, the client covers travel, accommodation and per diem expenses. Working and travel days, together with these expenses, are itemized and agreed in the proposal before work begins.
Pentesting, or penetration testing, checks whether a vulnerability can lead to data access or unauthorized actions. Oktsec runs controlled tests within the agreed scope and delivers findings, reproducible evidence and remediation priorities.
Yes. We test web applications, APIs, code, architecture and permissions. If your company also uses AI agents, we can include their workflows and tools in the scope.
Yes. We work with enterprises and startups in Argentina and Latin America, both before a launch and on systems already in operation.
Pricing and timing depend on the applications, APIs, repositories, workflows and environments in scope. We agree on scope and deliverables first; testing starts once you authorize the work.
A description of the system, workflows or repositories to test, their integrations and the reason for the assessment. Access and technical documentation are arranged later; do not send credentials through the form.
Yes. One retest of the original findings is included at no extra cost. The proposal defines which findings will be verified, the environment and the request window. New systems or changes outside the original scope are assessed separately.
Not necessarily. We can start in staging or a bounded environment. Code and system access are agreed for each assessment. Production testing requires explicit authorization.
No. A retest verifies fixes for the original findings. Continuous validation adds assessment cycles covering the agreed scope and its changes, to identify new risks and track unresolved findings. Frequency and terms are defined in the proposal.
Your team implements the fixes and we coordinate the retest. If you use AI agents, we can also assess runtime controls with Signal and Control.
Let’s talk about your environment
Tell us what you want to protect or which requirement you need to meet.
gus@oktsec.com ↗A platform beyond the pentest
Analyzes your agents’ tools and dependencies, powered by Aguara, our open-source engine.
Assessment ↗Assesses applications, APIs, code and agent workflows. Delivers findings, evidence and priorities.
Control ↗Enforces policies before connected tools run. Blocks actions, requests approval and records decisions.
Control acts on requests routed through configured integrations. The assessment does not require purchasing other products.