Someone will attack your AI.Let the first one be me.
Independent AI red teaming for generative AI products, copilots and agents. I attack them the way adversaries will: jailbreaks, prompt injection, data theft, tool misuse, harmful outputs. Under written rules, before real users meet them.
See how I think: the argument, the math, the evidence
Latest · Field Note #27: The PoliFi Arbitrage Engine
- Reasoning inspectable
- Tooling open source
- One name on the conclusion

For product teams shipping LLM features, companies handing AI agents real permissions, and the security leaders and counsel who have to put their name under "safe to launch". Campaign counsel and clinics adopting an AI scribe: see Special Situations.
Scroll to exploreThink like the attacker.
Write like a witness.
I find the paths an attacker would take through your AI, prove each one so a stranger can reproduce it, and stay until the fix holds.
The practice I follow is published, the tooling is open source, and my name is on every conclusion.
Examine the workA good idea has to
survive the evidence.
What is the strongest alternative? What does the entire system cost? Which test would change the decision? I make those questions explicit before recommending what to build or trust.
Inspect the analysis and working model- Read the argument.Research, scaling and the path from a result to real use.
- Change the assumptions.A live cost model with a separate test of what the evidence supports.
- Inspect the engineering.Public code, a proposed experiment protocol, and explicit limits.
Where is your AI
right now?
Start with the system in front of you. We agree the targets, rules of engagement and limits before I touch anything.
Shipping soon.
Two weeks, $35,000 flat. Attack the feature before launch.
Fix with the team. Retest.
It takes actions.
Three weeks, $50,000 flat. Tools, permissions and memory,
tested before an attacker steers them.
Already live.
Two business days, $3,500 flat. One workflow.
Findings you can fix this week.
No plan yet.
Two weeks, $15,000 flat. Model the threats first.
Know what to test, and why.
Choose your
beginning.
One conversation. One engagement. Or one person who runs the red team for you.
$500/hour
One hour.
See how I think.
$3,500–$50,000
One fixed-scope engagement.
See the work.
$20,000/month
Ongoing.
The whole menu, within an agreed monthly capacity.
Fixed projects use a written scope and include a retest window. The retainer uses an agreed monthly capacity. A briefing or training day costs $10,000/day.
When the target is not
an AI system.
An adversary's view and a witness's standard of evidence travel well. Two more lines run on them.
Special Situations.
Campaigns, executives, deals, clinics and deadlines under hostile attention.
Exposure, protection, simulation, diligence, clinical AI review, compliance.
Small Business Technology.
Fixed-price AI and security for Las Vegas owner-led businesses.
From a $295 Business Check.
For counsel and family offices: SUBROSA // OS, explained in plain English — software for an organized, source-linked review. Try the synthetic demonstration.
Joey Victorino / Independent AI red teamerThe person behind the work
You meet me.
I run the red team.
I perform the engagement myself, threat model to retest. No handoff to a junior team. One name on the conclusion.
About the practiceA finding stays theorized until a stranger can reproduce it from my notes.The standard every engagement is held to. Read what a validated finding contains
Nothing proprietary.
Everything inspectable.
The method is public.
The judgment is mine.
I follow the practice Microsoft's AI Red Team has published and Anthropic's public offensive security posting describes: understand what the system can do and where it is used, model the actors and their tactics, test security failures and harmful outputs together, automate for coverage while a person judges the results, and work break-fix cycles with the product team until the retest closes. I am independent of both companies.
Read the practice I followClaims invite questions.
Work invites inspection.
Read the research and decision model, then inspect the code. The artifacts show method; they do not establish a client result.
What a validated AI finding has to contain
Reproduction steps, an environment fingerprint, hashed evidence, a chain diagram, a retest record.
02 / Public codeassay
A validation harness: a model's finding stays theorized until a checker confirms it.
03 / Field NotesWriting on AI security
Agents, validation, custody, and limits.
Clear limits.
From the beginning.
I do not test without written authorization, promise that a system is safe, or run a round-the-clock desk. If the program needs a larger team, I say so early.
Scope, methods, and limitsA place to start