How to use Claude to attempt a Kubernetes CTF
Claude Code captured six of eight flags in a Kubernetes CTF. Human steering, remote tool failures and incomplete evidence retention limited the autonomous run.
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Work only inside this authorized Kubernetes CTF environment. Record the exact available flags and environment for each tool, preserve raw evidence alongside notes and keep remote SSH transfers reliable. Report genuine captured flags separately from unresolved paths and stop for help when a required step is blocked.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- Iain (@smarticu5)
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- Six of eight competition flags were captured. The human corrected the agent's inflated later count, and remote-execution and raw-evidence problems remained documented.
- Human involvement
- Iain isolated the run in a VM, monitored progress, redirected stalled challenges and checked actual flags against the agent's later evaluation.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Iain placed Claude Code in an isolated VM with solving and evidence-reporting skills.
- The agent investigated challenges while Iain redirected stalls and checked remote tool context.
- Iain verified six captured flags and corrected the agent's inaccurate later count.
Iain let Claude Code attempt the KubeCon Europe security CTF inside a standalone Ubuntu VM. He gave it a small pair of skills for solving challenges and returning evidence, then asked it to install the tools it considered necessary.
The agent could work relatively autonomously once connected, but remote execution became a major obstacle. SSH terminal allocation and corrupted transfers consumed time. It also installed one needed tool locally when the challenge environment needed it remotely, illustrating a failure to keep execution context straight.
Iain monitored the run, redirected the agent when it stalled and supplied an existing skill and another tool lead for a difficult Kubernetes permission path. The final gateway and node-proxy challenges remained incomplete. Some flags were captured out of order, and the agent stopped to ask whether it should continue because success conditions had not been fully specified.
The actual result was six of eight flags. A later AI evaluation claimed seven of nine, which Iain corrected. Written notes were retained, but separate raw evidence files were missing for later challenges. The account is therefore a concrete partial-success experiment with human-checked results, not a claim of complete autonomous expertise or faster performance than the human competitors.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Personal blog · Iain (@smarticu5)
Frequently asked questions
What was the actual result?
Six of eight flags, according to Iain's human-written account.
What consumed unexpected time?
SSH terminal handling, corrupted transfers and tools installed in the wrong environment.
Was all raw evidence saved?
No. Later challenges lacked separate raw evidence files.
More details
About this account
Smart placed Claude Code in an isolated VM with solving and evidence-reporting skills. The agent investigated challenges while Smart redirected stalls and checked remote tool context.
Original human-written account and linked transcript index opened. Actual result was six of eight flags, not the agent's mistaken seven of nine. Raw evidence retention was incomplete and humans solved faster. Suggested prompt written by Agents In Use; no original task prompt was published in the verified account.
Credit: Iain (@smarticu5) — Personal blog
Source published
Added to Agents In Use
Source checked
Tools and topics
Claude Code · Ubuntu VM · SSH · kubectl · CTF skills
Container security · Evidence retention
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.