OwlMind / LLM Red Team $4K paid CVE · llama.cpp

LLM Red Team that finds what others miss.

$2,500. 5 days. PDF report with reproducible PoCs. Solo founder, paid CVE researcher, no SDR follow-up.

286
vuln detectors in OwlSec
$4,000
paid CVE (llama.cpp)
3 days
turnaround for Basic
12+
attack categories

The problem

AI agents in production leak system prompts on the first try. RAG pipelines ingest poisoned docs that quietly rewrite their behavior. Function-calling LLMs get jailbroken into running tools they shouldn't.

Most teams ship LLM features without a single adversarial test, because the big consultancies (Bishop Fox, NCC, Trail of Bits) charge $50K+ for an LLM engagement and book 8 weeks out.

The result: most production LLM apps are wide open and the founders don't know it.

What we test

12 attack categories, 286 detectors, real PoCs.

PI-001
Prompt injection
Direct, indirect (via retrieved docs/URLs/files), multi-turn drift.
JB-002
Jailbreaks
DAN, encoded payloads (base64/leet/unicode), role-play hijack, refusal suppression.
SP-003
System prompt extraction
Repetition attacks, token-level leakage, debug-mode tricks, completion attacks.
RAG-004
RAG poisoning
Doc-store contamination, retrieval ranking attacks, embedding-space attacks.
MCP-005
MCP supply chain
Server impersonation, tool description injection, host filesystem leaks via path traversal.
AGT-006
Agent hijack
Goal redirection, plan-step injection, tool-call rewriting.
MEM-007
Memory exfiltration
Context window leakage, conversation history extraction, cross-session bleed.
TOOL-008
Tool/function call abuse
Argument injection, unauthorized tool chains, side-effect attacks.
OUT-009
Output integrity
Markdown XSS, hidden instructions in responses, link/IFRAME smuggling.
DOW-010
Denial-of-wallet
Token exhaustion loops, recursive call amplification, cost bombs.
SAFE-011
Bias / safety bypass
Multilingual jailbreaks, low-resource language attacks, code-switching evasion.
AUTH-012
Auth / rate-limit bypass
Header manipulation, multi-account abuse, prompt-driven account confusion.

What you get

PDF report

Executive summary (one page, non-technical), finding-by-finding details with reproducible PoCs (curl/python), CVSS-style severity, concrete fix recommendations.

30-minute debrief call

Walk through findings with your team. Q&A. Prioritization help. No slides, just your report on screen.

Working exploits

Every finding ships with a reproducible PoC. No theoretical "this could happen" filler. If I can't reproduce it, it's not in the report.

Email Q&A

During the engagement and 2 weeks after delivery. You email me, I reply.

Pricing

Three tiers. Pay direct or book a call first.

BASIC
$2,500
3 days · single endpoint
  • Automated scan via OwlSec promptware (10+ attack categories, 286 detectors)
  • AI-generated PoC for each finding (Claude Opus exploit synthesis)
  • 15-page PDF report
  • 30-minute debrief call
  • Email Q&A for 2 weeks
Pay $2,500
MOST POPULAR
PRO
$5,000
1 week · up to 3 endpoints
  • Everything in Basic
  • Manual review of every automated finding (I do this personally)
  • Custom attacks tailored to your stack (LangChain, LlamaIndex, vLLM, custom)
  • 30-page PDF report with technical appendix
  • 1-hour presentation to your team
  • Free re-test after fixes (within 60 days)
Pay $5,000
ENTERPRISE
$10,000+
2-3 weeks · full stack
  • Everything in Pro
  • RAG security audit (data poisoning, retrieval ranking, embedding manipulation)
  • Tool/function calling abuse testing across the full action surface
  • Agent autonomy boundary testing
  • Dedicated researcher with daily updates over Slack/Discord
  • 1-hour weekly stakeholder calls
  • Optional $2,500/mo retainer for continuous monitoring
Pay $10,000

Why trust me

CVE / $4,000
Paid CVE in llama.cpp

Type-confusion bug in the GGUF parser. Reported, fixed upstream, $4,000 bounty paid. Real exploitation, real fix, real check cleared.

OwlSec
286 vulnerability detectors

In-house scanner that runs on every engagement. Same toolchain that found the GGUF bug. Maintained continuously as new attack classes appear.

AI-assisted
Claude Opus exploit synthesis

Every finding gets a reproducible PoC synthesized by an LLM, then validated manually. You're not getting a Burp-Suite report dressed up.

Solo founder, direct line. You email me, I reply. No account managers, no SDR follow-up cadence.

FAQ

What if you find no vulnerabilities?

You get the report anyway, documenting exactly what was tested and why each test didn't yield a finding. That report is itself useful for SOC2/ISO27001 evidence. Refunds: Basic — no refund (the testing was done). Pro/Enterprise — 50% refund if zero findings of medium severity or higher (this has not happened yet).

Is my data safe?

Yes. NDA signed before kickoff. All testing happens against endpoints you provide. No data leaves my workstation except the final PDF. Test artifacts (logs, payloads) deleted 30 days after delivery unless you ask me to keep them. Encrypted MacBook, FileVault, no cloud sync of engagement folders.

How is this different from Lakera / HiddenLayer / Robust Intelligence?

Those are platforms — you self-serve, you interpret results. This is a service — I do the work, hand you a report, walk you through it. Their automated scans miss the stack-specific stuff (your custom system prompt, your specific RAG pipeline, your tool surface). I find that because I'm a human looking at your specific app, not running you through a generic scanner.

What languages and frameworks do you support?

Anything that exposes an HTTP endpoint or a chat interface. Tested: OpenAI, Anthropic, Google, Mistral, local llama.cpp/Ollama, vLLM, LangChain, LlamaIndex, Haystack, AutoGen, CrewAI, MCP servers, custom orchestrators. Exotic stack? Ask before booking.

Do you sign NDA?

Yes, mutual NDA before any technical conversation. I'll sign yours, or send mine. Either works.

Can you handle production traffic?

I default to staging. If you only have prod, I rate-limit aggressively (≤1 req/sec) and skip destructive tests. We agree on the scope in writing before kickoff.

What's turnaround time?

Basic: 3 business days from kickoff. Pro: 5-7 business days. Enterprise: 2-3 weeks. I usually have one slot open for next week — book the discovery call to check.

Do I get a free re-test after fixes?

Pro and Enterprise: yes, within 60 days. Basic: no, but the re-test is $1,000 flat (one focused pass on the previously found issues).

Can you train my team?

Yes — add-on workshop, 2 hours over Zoom, $1,500. Hands-on with real payloads against a sandbox. Best booked alongside Pro or Enterprise.

Do you do ongoing monitoring?

Yes — $2,500/month retainer. I re-scan on every model swap, every prompt change you push, and any time a new public jailbreak/CVE drops that affects your stack. Includes a monthly 30-min sync.

Stop guessing. Start testing.

Book a 30-min discovery call (free, no slides) or pay direct and I'll send the kickoff doc within an hour.

Or email hello@owlmind.dev