Skip to content

koscak.ai · cyber

Your last pentest
was a lie.
Your security team
is guessing.

The scanner ran. The PDF shipped. Nobody chained a single finding. You got a checklist, not an assessment. Our AI finds the average vulnerability in 23 seconds. The human operator behind it proves whether it actually matters.They ran a scanner, copy-pasted the output into a PDF, and charged you five figures. No one tested what happens when those findings are chained together. We do. And we start from the same scanner output, then chain.

AI doesn't care about your firewall.Your defenses were built for humans.

Three attack surfaces. AI-driven penetration. Human-verified chains. Every finding traced to blast radius.AI penetrates infrastructure that took years to build in minutes. Not because it is smarter. Because it tests thousands of paths simultaneously and never gets bored. Here is what that looks like pointed at your stack.

audit-diveAI penetration

23 seconds to first vuln.AI finds it in 23 seconds.

AI pre-scan identifies the average vulnerability in 23 seconds. Human operator chains it, verifies exploitability, maps blast radius. 23 minutes cold-start to first confirmed critical with full remediation path.Our AI hits your system and finds the first weakness in 23 seconds flat. Then a real human confirms it, chains it with other findings, and builds the actual attack path. The full report lands in 23 minutes. Your last audit took weeks.

See the methodologyWatch it work

ghost-huntAI vs. your AI

Your AI is an attack surface.That chatbot you shipped is wide open.

Prompt injection, RAG poisoning, tool-use abuse, model exfiltration, supply-chain compromise, perimeter erosion. Six classes tested against your production stack.Someone types the right sentence into your AI assistant and it dumps your customer database. Not a theory -- we do it in controlled tests every week. Prompt injection, data poisoning, model hijacking. Your chatbot has six holes your security team has never heard of.

See the six vectorsSee how it breaks

ironsightAI-mapped access

AI maps every trust boundary.We know who has access. You don't.

AI-driven graph analysis of cloud config, IAM policies, secrets, network topology. Every privilege edge proven or disproven. No manual spreadsheet audits.AI crawls your entire infrastructure and builds a map of every permission, every key, every forgotten test account that still has admin access. Three ex-employees can still log in. An API key from 2023 is in a public repo. You will find out from us or from an attacker.

See the permission graphFind out what is exposed

audit-diveDeep review

Outlandish precision.

Deep audit of a specific surface - web, API, cloud, model. Not a CVE list. A remediation roadmap.A focused review of one part of your system - website, API, cloud, or AI model. Not a long list of warnings. A plan for what to fix first, next, and eventually.

23min median first critical foundaverage time to find the first serious issue
koscak . audit-dive . example.com

chainedwalked through

IDOR → SSRF → RCEHow one small flaw opens the whole door

We don't stop at the first flag. Every finding is walked to its blast radius.We don't stop at the first problem. Every finding is traced all the way to how bad it could get.

mappedscored

CVSS 3.1 + CWE + OWASPIndustry-standard risk scoring

Auditor-ready tagging on every finding, not a generic severity estimate.Every finding is tagged using the scoring that auditors, insurers, and compliance teams actually recognise.

forensictamper-proof

SHA-256 evidence trailTamper-proof evidence trail

Every probe response hashed and archived. Reproducible, non-repudiable.Every step is cryptographically recorded so the findings are reproducible and can't be disputed later.

ghost-huntAI safety check

New stack.
New review.
New tech.
New questions.

LLMs, agents, RAG pipelines. Six classes of risk that weren't in your last pentest - worked through one at a time, in shadow-mode, against your stack.Your AI assistants, chatbots, and knowledge bases have new kinds of weak spots that older security reviews don't check for. We walk through six of them, safely, against your actual setup.

01

Prompt injectionHidden commands hijacking your AI

Attacker-crafted text hidden in documents, emails, or tool outputs hijacks your AI into running their instructions.Malicious text hidden inside documents, emails, or tool results that tricks your AI into following someone else's orders.

02

RAG poisoningTainted knowledge base

Seeding your knowledge base with adversarial content that changes what your assistant answers. Invisible in logs.Someone plants fake or misleading content in the knowledge your AI reads from - quietly changing the answers it gives, with no trace in the logs.

03

Supply-chain compromiseContaminated software parts

Malicious packages, poisoned weights, typosquatted deps. One bad import and you ship an attacker's backdoor.Bad code or bad model weights slip in through the software you install. One contaminated piece and you ship a back door without knowing.

04

Model extractionCloning your AI

Systematic querying that reconstructs your fine-tuned model's behaviour for the cost of an API budget.Someone uses your AI enough to rebuild a copy of it themselves - stealing the work you put into training it for the price of API calls.

05

Tool-use abuseAI running commands it shouldn't

Your AI agent has filesystem, shell, and network access. One crafted instruction and it runs attacker commands.Your AI can read files, run commands, and make network calls. A single crafted prompt can turn that access against you.

06

Perimeter erosionForgotten public endpoints

Dev envs exposed publicly. Webhooks from third parties. OAuth scopes past memory. Classic holes, modern scale.Test environments that accidentally ended up public. Old webhooks you forgot about. Permissions no one remembers granting. Classic mistakes at modern scale.

ironsightPermissions audit

Precision posture.

Every privilege edge mapped. Nothing guessed.Every permission mapped. Nothing assumed.

01

Cloud IAM edge-map.Who-can-do-what map.

Every role, every trust relationship, every permission boundary enumerated. We find the paths your compliance scan misses because it stops at the policy document.Every user, every role, every trust link - all written down. We find the shortcuts your compliance scan misses because it only reads the policy documents.

02

Network segmentation audit.Which machine can reach which.

Actual reachability tested from every segment, not just "the VLAN diagram says so." We ping every edge your engineers trust.We test what machines can actually reach each other, not just what the diagram says. The diagram is often wrong.

03

Secrets-in-source scan.Forgotten passwords check.

Full-history scan across every repo you own - not just current HEAD. The leaked key from 2022 is still valid until you rotate it.We search the full history of your code, not just the latest version. A password leaked in an old commit from 2022 still works until someone changes it.

04

Supply-chain pipeline.Imported code review.

Build steps, artifact provenance, dep-confusion surface. One bad import is the quiet win attackers are looking for.We check where your build pipeline pulls code from and whether anything unsigned or unexpected can slip in. A single bad import is usually all it takes.

hymnThe report

The report.
Built to be read.
The report.
Actually readable.

Executive summary for the board. CVSS-scored findings for security leads. Copy-pasteable remediation for the engineers who'll fix it. English always + any second language you request.A short summary for leadership. Risk-scored details for your security lead. Copy-paste fixes for the engineers who'll do the work. English always, plus any second language you ask for.

report_language

EN + ?

EN always included. Pick your second language. SK · CZ · DE · FR ship same-day. ES · PL · HU · UA · RO add 3 business days for linguistic review. Others on request.English is always included. Pick any second language you want. SK · CZ · DE · FR ship same day. ES · PL · HU · UA · RO add 3 business days for a review pass. Others on request.

12

findings / avg engagementissues per review (avg)

100%

CVSS + CWE + OWASP taggedindustry-standard scored

0

writes to your systemschanges to your systems

90 days

evidence retentionevidence kept, then deleted

compareCompare

Us vs. a scanner
vs. your in-house SOC.
Us vs. a scanner
vs. your own team.

Numbers, not narrative. One engagement, three lenses.Numbers, not marketing. One job, three ways of doing it.

koscak · audit-divekoscak · deep review

23m

to first verified criticalto first serious finding

Human-in-the-loop triage. Every flag hand-verified, chained into attack paths with CVSS and fix.A real engineer checks every finding by hand, connects the dots, scores the risk, and writes the fix.

Automated scannerOff-the-shelf scanner

14k

raw CVE matchesgeneric warnings

Qualys / Nessus / Burp Scanner. Fast, broad, noisy - and you still need to triage every finding yourself.Tools like Qualys, Nessus, Burp. Fast, broad, loud - and you still have to sort through every result yourself.

In-house SOCYour own security team

0

external-perimeter tests / yroutside reviews per year

Your team is busy running the business. An outside eye catches what familiarity hides.Your own team is busy running the business. An outside set of eyes catches what familiarity hides.

principlesHow we work

Quiet by default.
Independent by design.

A small practice. NDA before scope. Findings shared only with you. No conference talks, no case studies, no marketing reuse.A small team. We sign the NDA before we even talk about scope. Findings stay between you and us. No conference talks, no case studies, no marketing reuse.

01

Independent

no VC . no board . no quota

Privately operated. Every engagement is governed by the scope we agree with you, not a quarterly revenue target. We decline work we cannot deliver with outlandish precision.Privately owned. Every project is shaped by what we agree with you, not by a sales target. We turn down work we cannot do well.

100% privately operated
02

Deep

web . api . cloud . model . firmwarewebsite . api . cloud . AI . devices

We test the full stack. Every finding chained into an attack path, risk-scored, tagged to industry standards, mapped to controls. Remediation written by the engineer who found it.We test every layer. Every finding is traced to its worst-case, scored with industry standards, and fixed by the same engineer who found it.

5 stack layers tested
03

Quiet

NDA first . 90-day retention . 0 cloudNDA first . 90 days kept . 0 cloud

Findings never leave your engagement. No shared reporting portal, no cloud archive, no public write-ups. Evidence is encrypted, stored for 90 days, then shredded on confirmation.Findings never leave your project. No shared portal, no cloud archive, no public write-ups. Evidence is encrypted, kept for 90 days, then deleted on your confirmation.

0 cloud archive of findings

ghostrideSafe demo

See what's exposed
in 90 seconds.
See what's exposed
in 90 seconds.

Drop your URL. Non-destructive read-only ghostride. Lighthouse-style report of what's externally visible. No credit card, no trial limit.Drop your URL. We run a safe, read-only check. You get a report of what's visible from outside, in about 90 seconds. No credit card needed.

Non-destructive only. Read-only probes, no writes, no auth attempts. You get an instant scorecard.Safe and read-only. We never change anything on your site. Scorecard appears below in 3 seconds.

reviews

Quiet by default means we ship the work, then the stories.

First case studies land here with explicit customer permission. Until then, this slot stays empty. We will not fabricate a testimonial to fill a layout hole.

// 0 testimonials published · NDA holds

insights

Field notes.Field notes. When something breaks.

Short, opinionated, shipped the day something real happens. No fluff, no sponsored posts, no guessed takes.Short, opinionated, written the day something real happens. No fluff, no sponsored posts, no guessed takes.

View all posts ->

reading as
consumer pro