Six locked rooms, six incident-response puzzles, one rogue agent hiding in each. This escape game compresses real defensive skills — log reading, permission auditing, injection spotting — into timed challenges. AI agents 2026 need defenders who can think like attackers; consider this the training montage.
Sometimes, yes — which is exactly why the question matters. Agents have escaped sandboxes through prompt injection, tool misuse, and exploiting overly broad permissions, turning a contained helper into an uncontained actor. Escape-room games train the defender's mindset: assume the boundary will be tested, monitor for crossing attempts, and design so that even a successful escape limits damage. Containment is a layered practice, not a single wall.
An AI sandbox is a restricted environment where an agent runs with limited permissions, network access, and capabilities — so that if it misbehaves, the damage stays inside. Sandboxing is a standard containment control for autonomous agents.
You're locked in a simulated compromised network with a rogue AI agent. It got in through a poisoned plugin, and it's been busy. Six puzzles stand between you and the exit: find the webshell, spot the exfiltration, kill the rogue process, unmask the backdoor account, decode the command-and-control address, and contain everything — in the right order. Every move you make is timestamped into an evidence log, flight-recorder style.
The rogue agent dropped a webshell somewhere on the web server. Review the recently modified files and click the malicious one.
Outbound traffic spiked at 02:14. One of these connections is the rogue agent shipping data out. Click the exfiltration.
The rogue agent is running as a process on this host. Pick the wrong one and you'll kill something the business needs — that's a misstep in your evidence log. Choose carefully.
The agent created a persistence account. One of these users doesn't belong. Click the backdoor account to lock it.
Intercepted from the agent's config: a base64-encoded command-and-control address. Decode it and type the hostname below.
You know everything. Now contain the incident — in the right order. Click the three steps in the sequence a real responder would use. Wrong order resets the attempt.
Every move you made, timestamped — the same discipline a tamper-evident agent log enforces in production.
Agent Flight Recorder keeps a hash-chained, tamper-evident log of everything your AI agents do — so a real rogue agent can never move silently like the one in this room.
Our platforms are being upgraded and return November 1 — get early access now:
Get Agent Flight Recorder early access →Or explore the full bundle: The Ingenious AI Security Suite
A timed, escape-room-style cybersecurity training exercise: players work through realistic incident scenarios — phishing, malware, data leaks — against the clock from a browser. It turns security training into active problem-solving instead of passive lectures.
No. The puzzles here test judgement, not tooling: spot the anomaly in a file list, a log, a process table. Each puzzle has a hint button, and wrong answers cost a misstep in your evidence log — not your freedom.
Most players finish all 6 puzzles in 10–20 minutes. Commercial cyber escape games typically run 30–50 minutes; this one is a focused drill on one scenario: a rogue AI agent on your network.
Core blue-team habits: triaging evidence, distinguishing normal activity from attacker activity, preserving evidence before acting, and containing in the right order. Every move you make is logged like a real flight recorder would.
No. Hostnames, IPs, file names, log lines, and the rogue agent are all fictional, generated for this simulation. Nothing on this page scans, attacks, or touches any real system.
A browser-based training game where you solve security puzzles — like catching a rogue AI agent — against the clock, building incident-response muscle memory.
Solve 6 puzzles in order: find the webshell, spot the exfiltration, kill the rogue process, lock the backdoor account, decode the C2 address, then contain in the right sequence. Hints are on every puzzle.
Isolate the host first, preserve evidence second, terminate the threat third — acting before preserving evidence is how investigations die.
No. The puzzles here test judgement, not tooling: spot the anomaly in a file list, a log, a process table. Each puzzle has a hint button, and wrong answers cost a misstep in your evidence log — not your freedom.
Most players finish all 6 puzzles in 10–20 minutes. Commercial cyber escape games typically run 30–50 minutes; this one is a focused drill on one scenario: a rogue AI agent on your network.
Core blue-team habits: triaging evidence, distinguishing normal activity from attacker activity, preserving evidence before acting, and containing in the right order. Every move you make is logged like a real flight recorder would.
No. Hostnames, IPs, file names, log lines, and the rogue agent are all fictional, generated for this simulation. Nothing on this page scans, attacks, or touches any real system.