careers
AI Security Engineer
Secure systems that read untrusted text, call tools and act on the world.
An AI security engineer protects applications built on language models and agents: retrieval pipelines, tool calling, protocols such as MCP, and the evaluation that tells you whether they are safe to ship. The field is young and moves fast, so the job rewards fundamentals and careful measurement over memorized tricks.
see this path on the journey mapA day in the job
- Threat-model LLM features: what untrusted input reaches the model and what the model can do with it.
- Review retrieval and tool-calling designs for injection, data leakage and excessive agency.
- Design guardrails and permission boundaries around agents, and test that they hold.
- Build evaluation sets and regression tests so a prompt or model change cannot silently weaken safety.
- Assess third-party tools and servers that agents are allowed to use.
Stage 1· Foundations
Understand how language models, embeddings and chunking behave.
Study here
Learn elsewhere
- Python and calling a model API
- Basic machine learning vocabulary
- Prompt injection as a concept, and why it is hard to fully prevent
Build this
Build a tiny retrieval app, then write down every place where untrusted text reaches the model.
Stage 2· Practitioner
Secure retrieval, tool calling and the guardrails around them.
Study here
- Vector DB
- Retrieval
- RAG - retrieval-augmented generation
- Tool calling
- Guardrails - input and output filters
- Evals - quality measured in production
Learn elsewhere
- Access control applied to retrieved documents
- Structured outputs and schema validation
- Application security fundamentals for the surrounding service
Build this
Plant a malicious instruction in a document, watch your app obey it, then fix it with permissions and validation.
Stage 3· Advanced
Govern agents, third-party tools and quality over time.
Study here
- MCP - model context protocol
- Agent frameworks
- Orchestration
- LLM as judge - automated rubric
- Regression evals - fixed suite on the PR
- AI in triage and mitigation
- Securing the agent itself
- The verification loop
Learn elsewhere
- Supply-chain review of tools, plugins and servers
- Red-teaming methodology for AI systems
- Policy and risk management frameworks for AI
Build this
Build a regression suite from real failures and wire it as a gate that blocks prompt or model changes that make it worse.
What to show
- A threat model of an agent that can read email and call tools, with the permission boundaries you would impose.
- An evaluation suite for prompt-injection resistance that runs in CI.
- A review of a third-party tool server: what access it asks for, what could go wrong and what you would restrict.
Certifications (examples)
There is no mature, widely recognized certification yet: prefer public work such as OWASP LLM Top 10 exercises, written reviews and open evaluation suites.