0xreversed/with.ai
Research and benchmarksWorksSource available
Target: Flare-On 13 challenges

Ran Codex in Docker to solve Flare-On 13 reverse engineering challenges and placed 3rd

Flare-On 13 challenges: Codex in Docker took 3rd place in Flare-On 13.
Image: GitHub (mahaloz/ilio-mini)

What was done

A small harness runs Codex in a container against the Flare-On 13 challenges. The author reports a 3rd place finish, 80.5 minutes total and $21.3 in costs.

ilio-mini is a small harness by mahaloz that runs Codex inside Docker to solve Flare-On reverse engineering challenges. The repo says it placed 3rd in Flare-On 13, with a total time of 80.5 minutes and a cost of $21.3.

How it was done

The harness is a CTFd client plus a Docker launcher. Inside the container, a triage step and a Codex loop work through each challenge, then submit the flag and move on to the next one. The Codex config uses gpt-5.6-sol at high effort, with an AGENTS.md that holds the reverser's instructions.

Hooks keep the agent going until a flag is accepted. After 10 and 30 minutes without a flag, a nudge tells the root agent to split the work into lanes. Skills cover a Flare-On playbook, .NET, Python and Windows dynamic analysis. The agent also uses Kuna, an agent-based decompiler, which it called on 6 of the 9 challenges.

Results

The repo lists the 3rd place result, the time and the cost above. The author also published an analysis of tool use during the competition. Gaps the agent hit in Kuna are written to files so they can be fixed. This is a CTF setup, so the targets were challenge binaries made for the contest.