Building a Security Agent: Model choice, Harnesses and Evals
Picture: Web Directions

Conference In 5 weeks

December
Monday
to Tue 8 December 2026 · Sydney time

Building a Security Agent Model choice, Harnesses and Evals

Where
Research / Deep Tech room 488 George Street, 2000, Sydney
Host
Web Directions
Entry
Not stated
Sign-up
On the host's page
Open on webdirections.org

All Sydney events

You sign up with the host. foundercal only links to it.

Who you meetThe people on the listing, and what the host tells us about the room.

  • Simon Harloff Speaker
Investors Not stated The listing does not say
The format Talks and halls Conference
Sign-up On the host's page Price not stated

AboutIn the host's words.

Speakers: Simon Harloff (Dam Secure)

Track: Research / Deep Tech

In this presentation, I’ll show how we set out to give developers useful security feedback on every pull request in under three minutes. The benchmark results and methodology are published here: https://docs.damsecure.ai/blog/pr-review-security-benchmark-update/. It has become our most-cited research to date. We initially expected to compare models using established security benchmarks, but discovered that those tests made agents look more capable than they were. In our testing, frontier models recognised canonical vulnerable repositories such as OWASP Juice Shop, while markers intended for traditional scanners gave them clues that real pull requests would not contain. We needed private fixtures with known ground truth, realistic code and no familiar answers. We also needed to evaluate the complete review system, including the model, prompts, tools, context and stopping rules.

We first built synthetic replicas of open-source applications. They gave us controlled tests, but producing realistic codebases was too slow. We switched to reversing real CVE patches and replaying the vulnerabilities into matching open-source commit histories, then built a model- and harness-independent eval suite that records findings, files opened, searches, tool calls, runtime and cost. I’ll demonstrate the approach by running a synthetic pull request containing an IDOR through an open-source harness, revealing the hidden ground truth and inspecting the agent’s trace. Our benchmark covered ten planted access-control bugs, with five runs per model. Recall ranged from 30 to 100 percent and cost ranged from about $0.04 to $4.25 per pull request. I’ll close with what we would change: start with replayable workloads, capture traces from the first run, and choose metrics based on the failures developers care about.

https://data.webdirections.org/ai-engineer-sydney/sessions.json#55fb0cbc-f032-4c0a-840a-f53cc92b6372

Full text on webdirections.org .

Where

Research / Deep Tech room

488 George Street, 2000, Sydney

Open in Maps ↗

Same week in Sydney All 151 events

Dec
to 5 Dec
Hackathon

AI Engineer Hackathon

Founders pitch

Web Directions Free
Dec
Fri 08:00
Dinner

Startup Coffee Social #39

Investors in the room

Tank Stream Suites
Startup Coffee Social Events
A$5
Dec
to 13 Dec
Conference

AI Week

Web Directions
Dec
to 12 Dec
Conference

NeurIPS 2026

One of the year's big ones

NeurIPS Foundation
Dec
to 8 Dec
Meetup

Lessons from Economics for the Human / Agent Software Workforce

Leadership room
Web Directions
Dec
to 8 Dec
Meetup

Ontologies

AI Engineering room
Web Directions
Dec
to 8 Dec
Meetup

Shipping together in an AI native team

Software Engineering room
Web Directions
Dec
to 8 Dec
Meetup

Dependency hell is back. This time it's your agent's config.

Software Engineering room
Web Directions
Dec
to 8 Dec
Meetup

Your website is a terrible API

AI Engineering room
Web Directions
Dec
to 8 Dec
Meetup

Sovereignty Is an Inference Problem

Leadership room
Web Directions
Open on webdirections.org Mon 7 Dec ↗