AI Red-Team / Safety Evaluation Engineer
Open to mid-level and senior IC roles, W2, remote (US).
I build automated tests that try to break AI systems, and record which attacks get through.
changeadversarial probeattempt caughtgatereceipt
Requirements and proof
| Requirement | Proof |
|---|---|
| Design test methods that probe an AI or agent system for security weaknesses | Adversarial red-team suite - Public adversarial-gate repo |
| Encode security checks as a reproducible, automated pipeline instead of a one-off manual review | Release-gate case study - Deterministic gate battery case study |
| Combine deterministic checks with model-assisted analysis while holding false positives near zero | Deterministic gate battery case study |
| Build detection tooling that catches an attempt end to end, not just an alert on it | Adversarial red-team suite |
| Keep a re-runnable record of what was tested and what passed or failed | Governance-mapping case study |
| Reduce manual security review through code-based, policy-as-code automation | Release-gate case study |
What could go wrong hiring me
Most proof here is self-built, not earned at a household-name employer.
Counter: Public adversarial-gate repo with CI
The attack classes tested are the classes I chose to test, not an independent list.
Counter: OWASP-mapped red-team case study
This work has not been through a live bug-bounty or third-party penetration test.
Verify in 60 seconds
- Clone the public adversarial-gate repo and run: pytest llm-adversarial-gate/ -v, https://github.com/mpuodziukas-labs/llm-adversarial-gate
- Open the red-team case study and check the attack classes against the assertion count, /cases/adversarial-red-team
- Open the deterministic gate battery case study and read which false-green claim it rejected, /cases/deterministic-gate-battery