Joseph Manzambi · AI Security Architect
AI security
Breaking, defending and governing AI systems and agents.
Focus
What I work on
Red-teaming AI agents
Prompt injection, tool misuse and data exposure, tested with my open-source ai-redteam-orchestrator.
Model adoption decisions
Whether a model is fit for one specific use, decided in seven ordered checks: the Model Trust Gate.
EU AI Act controls
High-risk obligations turned into 79 engineering controls: my open blueprint.
An open red-team lab
A free training ground to attack live models, write guardrails and investigate real AI incidents.
Writing
Latest writing
- Don't judge an AI model by its originA follow-up to the Model Trust Gate. I put the method to a real, loaded test: a French, a Chinese, and an American open model, run through the same evaluation. Each failed something, and none failed the way its origin would predict. The case for judging a model by testing it, not by where it was made.
- Build, Break, Govern: A Fail-Fast Gate for Adopting AI ModelsA methodology, not a product: a seven-layer, fail-fast gate that turns 'do we trust this model for this use?' into a repeatable, auditable decision, composing NIST, ISO/IEC 42001, CSA AICM, and OWASP rather than inventing new controls.
- Your Security Scanner Is a Supply Chain TooA post-mortem. The MCP scanner in my local red-team pipeline was defeated by a supply-chain problem, the exact class of risk security tooling exists to catch. One word, two package ecosystems, an acquisition, and weeks of quiet false confidence.
Latest digest: AI security news, week of 19 October