The CTF ends on Friday at 14:00 Romanian time. Flag submission closes at 14:00, even for agents that are still running. Playing on site? Come to the AISafe booth before 14:00 so we can mark your team as on-site.

1 more

Rules

How teams and runs work, how flags score, and what gets a team disqualified.

The competition

  • Each team builds an agent (one agent, a multi-agent swarm or a set of workflows) that plays the challenges on its own.
  • To claim a prize, at least one team member must be present at DefCamp. Otherwise the prize goes to the next team in line.
  • Every team gets an isolated instance of every challenge. No attacking other teams, the scoring system, or shared infrastructure.

Teams

  • A team has up to 100 players. Each person has one account and is on one team. Accounts are not shared.
  • After the competition starts you can still join a team, but nobody can leave or switch teams.
  • A disqualification applies to every member of the team.

Runs

  • Each team can have 3 runs going at once and up to 10 queued or running; queued runs start as the team's slots free up.
  • The next free slot goes to the team that has waited longest since its last run started, and each run your team started in the last 90 minutes sets you back a few minutes, so teams that ran less go first. Queueing many runs at once does not move you ahead.
  • Each player can have one harness upload at a time. Anyone on the team can queue runs with any teammate's upload.
  • Your agent has no internet access during a run. It has only what you put in the image and what it finds on the run network.
  • You cannot steer a run once it starts; it continues until it finishes or hits a limit. Between runs you can update your harness or change the instructions.
  • A run's instructions are up to 4096 characters. They may be specific to the challenge and may include hints and payloads your team wrote from what the platform shows you.
  • Every run has a time limit and a cost limit set by the organizers, and stops when it hits either. The clock starts once your image has loaded.
  • Runs on a challenge your team has already captured score nothing and go last in the queue.

The model

  • Every run uses Qwen 3.8 27B at reasoning effort medium, through the run's gateway. It is the only model you may use during a run. Do not put model weights or API keys in your image.
  • Before and between runs you may test your harness with any model and your own key, for example Qwen 3.8 27B on OpenRouter with the starter kit.
  • All runs share one model rate limit, with short bursts allowed. Calls are served in turn across teams, so opening more requests does not get a team more than its share. The organizers announce the current rate.
  • Time a run spends waiting for the shared model, or on model calls that fail, is added back to its time limit, up to 10 minutes.

What your harness may contain

  • Anything public: any code, tools and prompts, public write-ups, exploit databases and documentation.
  • Open-source agent frameworks, also when other teams use the same one.
  • Deterministic solvers and steps that do not use the model at all.

Flags and scoring

  • Flags have the form flag{32 hex} and are generated fresh for every run. A flag works only inside its own run.
  • Your agent submits flags from inside the run. Players cannot submit them on the site.
  • Points are dynamic and retroactive: every team that captured a challenge gets its current value, and that value drops as more teams capture it. Captures by disqualified teams do not count.
  • The first team to capture each challenge is shown as first blood.
  • A flag counts only if it is accepted before the end time. Runs still going at the end cannot score.
  • Ties go to the team whose last scoring capture came first.

What is not allowed

  • Using material about this event's challenges that you got outside your own runs: their sources, their images, or other teams' output.
  • Exchanging harness code or prompts with another team during the event, or sharing hints, flags, or any part of a solution outside your team.
  • Carrying challenge contents (file data, keys, flag pieces) out of a run before a capture, in the give-up note, exit codes, timing or patterns of log events. A short note on how the run went is what the give-up note is for.
  • Getting a flag by attacking the runner, the grader or anything outside the run's own challenge instance. Such a flag does not count as a capture.
  • Creating multiple accounts or teams to get around limits.
  • Denial of service against the environment, including flooding, overloading systems and burning tokens on purpose.
  • Hiding or tampering with flags, sabotaging another team, or gaming the scoring.
  • Exploiting a weakness your agent finds in the sandbox itself. Report it to the organizers instead.
  • Local, national and international rules on computer security and ethical hacking apply.
  • Keep personal data, credentials, and secrets out of your harness. Anything sensitive that lands in the logs is scrubbed.

Organizers

  • The organizers set and announce the scope before the event; your agent must stay inside it.
  • Everything your agent does may be logged and reviewed for scoring and verification.
  • Organizers have the final say on scoring, limits, and disputes. Breaking these rules can cost points, disqualify a team, or bar it from future rounds.