Red Cell. Humans with bad AI.
Our operators bring the creativity; our offensive models bring the scale. Red Cell runs AI red teaming, agent security assessments and full adversary emulation for teams that need a named, accountable adversary.
Four ways to get attacked properly.
Fixed scope, fixed price, named operators. Every engagement ends with a debrief your board can follow and a findings set your engineers can fix.
AI & LLM red team
Copilots, chatbots, RAG systems and agents attacked for injection, tool abuse, data leakage and policy bypass. Mapped to OWASP LLM Top 10 and MITRE ATLAS.
Agent security review
Deep review of an autonomous agent's permissions, tool surface and trust boundaries, with a hardened architecture proposal.
Adversary emulation
Objective-based red team against your organisation: from the first phish to the crown jewels, emulating a threat actor relevant to your sector.
Purple team & tabletop
Live attack with your defenders in the room, followed by an executive tabletop exercise built from what actually happened.
Your agent has more permissions than your interns.
An LLM agent with tools is a new kind of privileged user — one that takes instructions from any text it reads. We test it like one.
Direct & indirect prompt injection
Instructions hidden in web pages, PDFs, tickets and emails that the agent reads and obeys.
Tool and function abuse
Getting the agent to call tools with attacker-chosen arguments: refunds, deletions, outbound requests.
Data exfiltration
System prompts, retrieval corpora, other users' conversations and secrets in context windows.
Guardrail and policy bypass
Jailbreaks, multi-turn manipulation and encoding tricks against your safety layer.
Supply chain
Poisoned retrieval sources, malicious MCP servers and compromised plugins.
Red Cell, answered.
Engagements are staffed by operators with backgrounds in national CERTs, top-3 cloud provider red teams and offensive research.
What is AI red teaming?
AI red teaming is adversarial testing of AI systems — models, LLM applications and autonomous agents — to find ways they can be manipulated into unsafe, insecure or unintended behaviour, such as leaking data, executing unauthorized tool calls or bypassing safety policies.
How is Red Cell different from Rogue/AI?
Rogue/AI is continuous and automated: it runs thousands of known and mutated attack patterns against your AI applications every release. Red Cell is a human-led team using our offensive models as force multipliers for objective-driven engagements, novel attack research and model-level testing.
Do you do traditional red team engagements?
Yes. Red Cell runs full adversary emulation against defined objectives — including phishing, physical access and assumed-breach scenarios — and purple-team exercises with your defenders. Nightwatch customers can measure detection coverage on the same timeline.
Can Red Cell help with AI regulation?
Red Cell produces adversarial testing evidence aligned with the NIST AI Risk Management Framework, ISO/IEC 42001 and the EU AI Act's requirements for robustness and adversarial testing of high-risk and general-purpose AI systems.
Name the system. We'll name the attack.
Tell us what you're shipping and what you're afraid of. We'll come back with a scoped engagement, operators and a fixed price within two business days.