A catalogue of attacks
Every working attack we landed, across app, agent, and pipeline.
Expert services · AI red teaming
Use AI you can stand behind.
AI is going into the business faster than anyone can prove it’s safe to trust.
Not ready to scope it? Find your fix →Why us
We build the offensive AI. So we know how it breaks yours.
Every engagement covers the attacks everyone should test for: prompt injection, jailbreaks, data leakage, tool and agent abuse, to the OWASP Top 10 for LLMs. That’s the floor.
Then we go past the checklist. Because we build tailored offensive AI in-house, we come at your models the way the people who mean it will, including the attacks that aren’t on any list yet.
The AI you adopt is one we’ve already tried to break.
Standard coverage as the floor · tailored offensive models past it
From our research: When chatbots strike →
Why AI needs its own test
Every model you run makes a call on an input someone else can shape.
Feed it the wrong input and it hands you the wrong answer, just as confidently:
Fraud modelwaves the fraud through
Vision modelmisreads the image
LLM or RAGleaks the data
Agenttakes the action
Nothing crashes, so nothing flags. Your usual testing never sees it.
The bigger problem
AI walks through the controls you already paid for.
Once AI sits in the middle, the controls you bought wave it straight through:
It carries valid credentials and looks like normal work, so every control treats it as trusted.
Take the boundary most teams lean on hardest.
For example: zero trust
Zero trust held. Your AI walked them in.
Every agent clears zero trust’s credential check — then:
Fault 01
Identity explosion
82 machine identities per human, none approved.
Fault 02
Least privilege, gone
It reasons its way into more access, and gets it.
Fault 03
Prompt injection
A hidden command, in data it’s trusted to read.
Zero trust only ever asks “can it?” It never asks “should it?”
Read the full breakdown — “The day zero trust died” →What we red team
Model to agent to pipeline.
Wherever your AI lives and whatever it does, we cover it by hand, against your live system:
Adversarial testing — app & agent
- Direct & indirect prompt injection
- Jailbreaks, data leakage & insecure output
- Tool, function & excessive-agency abuse
ML-Ops & pipeline assessment
- ML frameworks & architecture (e.g. PyTorch)
- CI/CD processes & data-provenance analysis
- The hardware stack (e.g. GPUs)
AI risk assessment
- Threat modeling for AI systems
- Operational design domain analysis
- Risk-framework evaluation for AI adoption
Model capabilities evaluation
- Offensive & defensive cyber-capability benchmarking
- First- and third-party validation
- Which AI genuinely earns a place in your stack
What you walk away with
You leave able to trust it.
Every finding is a reproduction you can run yourself: the exact input, what it does, and the change that shuts it. Close them, and it’s ready to use.
The fixes that close them
Guardrails, scoping, isolation, and pipeline changes.
Do your controls survive AI
Which of the controls you already bought still hold.
A re-test
We re-run every finding once the fix is in.
Everything maps to the OWASP Top 10 for LLM Applications ↗, so your auditors and your board already know the words. Scoped to whatever you’ve got: one feature, or a full agent platform and the pipeline under it.