Tolmo · posted · checked
Research Engineer, AI Security
PythonTypeScriptOn-site · San Francisco
Who can apply
On-site · San Francisco
Location: San Francisco
ABOUT TOLMO
Tolmo is cybersecurity for the age of AI. Coding agents now ship software faster than any human team can review or secure it; 100+ pull requests per engineer per week is no longer an outlier, and attackers are using the same agents to find and exploit weaknesses in days, not months. The old playbook of quarterly pentests and endless alert queues breaks at this speed.
We're building a new security playbook. Tolmo runs a fleet of AI security agents that reason over a live graph of your entire production system (code, CI, cloud, and telemetry) and continuously attack a faithful copy of it like a real adversary. Every finding arrives validated, with its blast radius and a remediation plan that lands in your workflow, ready for a coding agent to fix. Hours to integrate, days to your first validated finding.
We've done this before: Tolmo's founders built Sqreen (acquired by Datadog) and led security products that thousands of organizations trusted. We're backed by Accel and Y Combinator and just getting started.
THE ROLE
Our product is only as good as our agents' judgment. As a Research Engineer on AI Security, you make that judgment better and measurably so: how our agents plan, reason over a graph of a customer's production system, decide what is worth testing, and know when a finding is real.
This sits between research and engineering. In one week you might design an eval that catches a new class of false positive, change how an agent plans a multi-step investigation, compare models on a hard benchmark, and harden the agents themselves against prompt injection from the systems they inspect. You'll work directly with the founders, our engineers, and the security leads who run pentests with our agents, and your work ships to production.
WHAT YOU'LL DO
Agent reasoning: improve how our agents plan, use tools, and reason over a live graph of code, cloud, and telemetry, so they reach real findings faster and with fewer dead ends.
Evaluation: build the evals, benchmarks, and graders that tell us whether an agent actually got better, and make them part of how every change ships.
Model selection and tuning: compare models, prompts, and agent designs on cost, latency, and quality, and turn the results into production decisions.
Model and agent security: defend our agents against prompt injection, data exfiltration, and manipulation by the very systems they test, and research the security of LLM-based applications our customers build.