Bay Area · Trailhead, 4,035 ft
Farzaan Kaiyom
Head of AI at Corridor, building the security layer for AI coding agents. Previously Vooma and Stanford CRFM.
01Selected work
Corridor · 2026
Argos: Corridor's first security model
A fine-tuned Qwen3.5-9B that reads a coding agent's plan and picks the security guidance it needs before any code is written. It posted the highest micro F1 of 14 models on our SecReT benchmark and the fastest time to first token. Built with Aditi Narasimhan; we presented it at DEF CON 34's AI Village.
Transluce · 2026
AI agents targeting government websites
With Transluce, MIT, and AIUC, we traced rogue AI agents probing U.S. and Canadian government sites: two failed hacking attempts, against the U.S. Department of Education and Library and Archives Canada, plus a wider pattern of aggressive tactics for pulling public data. None reached anything that wasn't already public.
Corridor · Current work
The security layer for AI coding agents
Corridor reaches coding agents like Claude Code, Cursor, and Copilot through an MCP server and agent hooks, reviewing an agent's plan before any code is written and then reviewing the pull request it opens. I lead the AI work behind it.
ICLR 2025
AutoBencher
Benchmark construction as an optimization problem: declare what you want (difficulty, salience, novelty) and let a language model search for datasets that satisfy it. The result elicits 22% more model errors than the benchmarks it replaces.
Stanford CRFM
HELM Safety
Safety evaluations usually run behind closed doors, which leaves the public unable to interpret them. HELM Safety is the public alternative: standardized benchmarks across six risk categories, run over two dozen leading models.
Vooma · Founding engineer
An AI workforce for logistics
Automating the quoting, order entry, and document handling that freight brokers used to run by hand. Customers include Echo, MoLo, MODE, Sunset, Evans, and NFI.
02About
At Corridor we build guardrails that reach a coding agent while it's still planning, review the pull requests it opens, and watch everything it writes. Most of my time goes to agent security evaluation: red-teaming coding agents, building the benchmarks that show where their blind spots are, and training models like Argos to close them.
I got here from both sides of the problem. I was a founding engineer at Vooma, a machine learning engineer at PayJoy, and co-founder / CTO of Gradia Health (YC W21); before that, a research engineer at Stanford CRFM, working on HELM and publishing on benchmark construction and safety evaluation.
03Get in touch
Always up for a conversation about AI safety, agent security, or evals. And we're hiring at Corridor.