Collinear AI’s Blog

Collinear Zero: a Pareto-optimal defensive cyber agent

TL;DR We built Collinear Zero, a harness for defensive cybersecurity agents.

21 hrs ago • Arthur Chen

17

Unlearnability in RLVR: What More Samples Reveal About Hard Tasks

We study unlearnability in reinforcement learning (RL) tasks for language models.

Oct 27 • Chinmaya Andukuri and Guga Gogia

13

1

Introducing MatFlowBench

Measuring whether AI agents can carry a materials question to an answer the evidence supports.

Oct 9 • Erekle Jmukhadze and Guga Gogia

24

1

1

September 2026

CWE-bench v1: Defense Is the Harder Test

Harder to game. Fairer to measure.

Sep 28 • Gonzalo Gonzalez, Ajay Jagannath, Adit Jain, and Nazneen Rajani

8

The Measure of Hard but Fair

A benchmark task can defeat a capable system and still tell us little about how systems differ. Three tasks from our security evaluation show what pass…

Sep 11 • Guga Gogia

18

CWE-bench: Measuring Coding Agent Capabilities on Every Class of Known Vulnerability

CWE-bench is a frontier benchmark for AI Cybersecurity

Sep 2 • Gonzalo Gonzalez, Adit Jain, and Nazneen Rajani

19

August 2026

The Vulnerability in the Reward

When the score is also the reward

Aug 28 • Guga Gogia

15

The Simulated World Is Being Born In Cybersecurity

AGI needs simulated worlds that are hard enough to expose failure and fair enough to make that failure useful. Cybersecurity is forcing these worlds to…

Aug 25 • Nazneen Rajani

13

The Uninterpretable User

What Simulated Users Say When the Agent Isn’t Listening

Aug 6 • Parker Seegmiller, Chinmaya Andukuri, and Guga Gogia

12

July 2026

The User Went for a Cigarette

Modeling Partial Observability for High-Fidelity User Simulation

Jul 1 • Parker Seegmiller, Chinmaya Andukuri, and Guga Gogia

10

May 2026

Is your RL environment fair to your agent?

or ensuring that your hillclimbing budget is spent right :)

May 14 • Adit Jain

9

Whose Taste?

More data won't fix the AI verification problem. Different taste might.

May 7 • Sachin

15