AI Red Team Engineer

Hamming AI•Austin, TX

4d•Remote

About The Position

Hamming is hiring an AI Red Team Engineer to build and own the harness that repeatedly tests voice and chat agents against authorized adversarial scenarios. The role turns techniques identified by researchers into repeatable tests run at scale, with evidence customers can inspect and results that can be replayed across releases. The engineer will build the runner for calls and chat sessions, evaluators that interpret outcomes, and replay and comparison capabilities. The system should support thousands of concurrent calls while remaining bounded, cancellable, and observable. The role also involves generating useful scenario variations, distinguishing successful attacks from refusals, infrastructure failures, and inconclusive results, and converting validated findings into regression tests that run on every change. Hamming describes its working style as customer-focused, fast-moving, and evidence-driven, with daily shipping and ownership of results. The role may be remote within the United States for most positions, with offices in San Francisco and Austin.

Requirements

At least 3 years of experience building backend services, asynchronous workers, or test infrastructure in Python or TypeScript, including handling failure cases and writing tests.
An adversarial mindset demonstrated through offensive tooling, CTFs, fuzzing, or a security project; a specific job title is not required.
Good judgment about evidence, including knowing when automated evaluation is sufficient and when human review is needed.
Experience or ability with idempotency, backpressure, observability, and enforcing limits in systems.

Nice To Haves

Experience with agent frameworks, voice or telephony systems, or LLM evaluation infrastructure.
A security finding of your own that was responsibly disclosed.

Responsibilities

Build the attack harness, including the runner for calls and chat sessions, evaluators, and replay capabilities.
Turn new attack techniques into definitions within the existing testing system rather than building a new system for each technique.
Generate useful variations by branching multi-turn scenarios and voice conditions with explicit objectives and conditions.
Make results trustworthy and inspectable by distinguishing successful attacks, refusals, infrastructure failures, and inconclusive results.
Make findings reproducible by capturing scenarios, configuration, model versions, tool events, outcomes, replays, and comparisons.
Convert validated findings into regression tests that run on every change while controlling noise.
Ensure the harness can handle thousands of concurrent calls while remaining bounded, cancellable, and observable.