- Run end-to-end attacks against LLMs, agents, RAG pipelines, and surrounding applications.
- Turn jailbreaks and unusual model behavior into reproducible exploits with evidence of impact.
- Trace attack paths through tools, credentials, identity, cloud, and data to establish blast radius.
- Exploit tool and function calling, MCP servers, plugins, memory, and multi-agent handoffs.
- Attack AI supply-chain components, including training data, embeddings, model artifacts, and MLOps infrastructure.
- Review agent frameworks, orchestration logic, and model-serving code for vulnerabilities.
- Use black-box probing, adversarial inputs, model extraction, and membership inference to assess model behavior.
- Translate AI attack tradecraft into techniques, decision logic, and evaluation data for agents.
- Pressure-test agent models and make human-in-the-loop safety and accuracy decisions.
- Build scripts, harnesses, and tools to automate AI attacks, and report attack narratives, business risks, and remediation.
PythonBash