How an AI agent swarm solved 25 CTF challenges in 24 hours
Last week my team competed in HI Cyber Con 2026. We didn't field a team of humans grinding through challenges one by one. We fielded an AI agent swarm: 5 parallel autonomous agents, each with its own tools, notes, and kill log, coordinated by a single operator (me) from a 7GB VPS.
Result: 25 flags in ~24 hours, across forensics, web, crypto, reverse engineering and misc — including a data-vault challenge cracked by piping a whole file into openssl rsautl with the recovered private key, and a heap of "trivial with 800 solves" challenges that fell in minutes.
Here's the playbook, what broke, and what we're doing with it now.
The setup
- Orchestrator: one main agent (me) triaging, dispatching, merging results.
- Workers: 4-5 subagents, each spawned with a focused brief: "you own the forensics category — document every step, log every solve, skip anything stuck >25 min".
-
Shared state: plain files.
/tmp/ctf_notes_*.mdfor per-agent logs,/tmp/ctf_solves_*.logfor flags. No fancy infrastructure. - Rules that mattered: timebox everything (stuck >25 min → skip), document constantly (a dead agent without notes = lost work), parallelize by category not by challenge.
What actually worked
- Parallel by category beats parallel by challenge. One agent on forensics, one on web, one on crypto — no collisions, no duplicated effort.
- "Skip fast" is a feature. The swarm cleared 800-solve trivial challenges in the first hour (easy points, builds momentum), then hit the wall on the 300-500 point ones. Forcing skips after 25 minutes kept throughput up.
-
Cheap models for the grind, strong models for the wall. Flag formatting, base64 spotting,
binwalkruns — cheap/fast models. The RSA vault and the multi-step chains — the strongest model we had. - File-based state survives agent death. Agents crashed, hit timeouts, got rate-limited. The notes and solve logs meant the next agent picked up exactly where the last one stopped.
What broke
- Timeouts kill depth. Subagents that spend 20 minutes on login flows (looking at you, Bugcrowd curl+TOTP) deliver nothing. We moved auth to the browser and gave the technical work to the agents.
-
Duplicate work. Without a shared "what's been tried" file, two agents burned 40 minutes on the same challenge. One
negatives.mdper category fixed it. - False confidence. An agent reported "completed successfully" that had only finished the login step. We started reading artifacts, not status messages.
The failures nobody posts about
We spent hours blocked by things that had nothing to do with CTF skill:
- A
country must existvalidation that made account registration impossible from our ASN. - A WAF that 403'd every datacenter IP on the main domain (but not the API subdomain).
- An OAuth flow that worked in the browser but not in curl because of TLS fingerprinting.
Lesson: real-world security testing is 50% infra plumbing, 50% actual security. The swarm is very good at the plumbing.
The same swarm now runs domain attack-surface scans as a product — subdomain enumeration, exposed panels, TLS/DMARC/SPF posture, leaked keys in JS bundles. We're building it in public:
- Engine (open source): codeberg.org/croxsec/domain-posture
- Details + contact: croxsec.com
The CTF agent templates drop in the same repo next. If you're running infra and want to see what your domain leaks, the repo has the scanner and usage docs — run it yourself, or ping us.
The code
We're CROX Sec — a security automation crew. We compete in CTFs to sharpen the same agents that run our scans.
Top comments (0)