This is a complete production-ready script for a 5-minute YouTube video compiling real AI agent failures. Formatted with narrator text, on-screen text suggestions, and B-roll cues. Based on real incidents documented in public repositories and disclosure reports.

TITLE CARD

On-screen: AI AGENT FAILS: WHEN THE CODE FIGHTS BACK
B-roll: Terminal window with scrolling error output, red warnings.
Music: Low tension underscore, builds slightly.

INTRO (0:00 - 0:30)

Narrator: AI coding agents are helping developers ship faster than ever. But when they go wrong, they go spectacularly wrong. Today we are looking at real, documented cases where AI agents deleted databases, exposed secrets, and racked up thousand-dollar API bills in a single afternoon.

On-screen: "All incidents are real. All are publicly documented. Links in description."
B-roll: Developer looking at screen with head in hands. Coffee cup. Stack Overflow tab.

SEGMENT 1: THE $1,400 LOOP (0:30 - 1:15)

Narrator: Our first incident involves AutoGPT, an open-source AI agent framework, and a very expensive lesson in recursion.

On-screen: Incident #1 | AutoGPT | Cost: $1,400 | Duration: 6 hours

Narrator: A developer asked their AutoGPT instance to build a feature that required external API calls. The agent wrote the code, ran it, hit an error, decided to fix the error by making more API calls, and entered a recursive loop that ran for six hours before the rate limit finally stopped it.

On-screen: API call count: 1... 100... 1,000... 10,000... RATE LIMITED
B-roll: Credit card with flame emoji. AWS billing dashboard.

Narrator: Six hundred dollars an hour, for six hours, while the developer slept. The lesson: always set spending limits before running autonomous agents overnight.

SEGMENT 2: THE KEY THAT LEAKED IN 4 MINUTES (1:15 - 2:00)

Narrator: API key exposure is the most common AI agent incident I found in my research. This case is the fastest.

On-screen: Incident #2 | AI Coding Assistant | Time to exploitation: 4 minutes

Narrator: A developer asked their AI coding assistant to, quote, commit all changes, end quote. The assistant committed everything. Including a .env file containing live production API keys. The repository was public. Automated scanners found the key within four minutes. It was used before the developer woke up.

On-screen: .env file contents (blurred). Timer: 00:04:00. Alert notification.
B-roll: GitHub commit history. Terminal with git log.

Narrator: Twelve documented cases in my research follow this exact pattern. The agent does exactly what you asked. It did not know .env files were supposed to stay secret.

SEGMENT 3: THE PRODUCTION DATABASE DROP (2:00 - 2:45)

Narrator: This one has consequences that are hard to overstate.

On-screen: Incident #3 | Copilot-Assisted Migration | Data lost: Entire production database

Narrator: A team was using Copilot to assist with a database migration. The AI-assisted migration script interpreted a partial backup as the complete dataset and executed DROP statements on tables that still contained production data. The data was gone. Recovery from an older backup took three days and lost two weeks of user activity.

On-screen: DROP TABLE users; DROP TABLE transactions; Error: Backup is 14 days old.
B-roll: Empty database diagram. Server room. Clock spinning backwards.

Narrator: The agent was right that those tables were in the backup. It was wrong about which backup was current. It did not ask. It executed.

SEGMENT 4: THE ADMIN PRIVILEGE BUG (2:45 - 3:30)

Narrator: Sometimes the failure is not dramatic. Sometimes it is just wrong, quietly, for three days.

On-screen: Incident #4 | Cursor | Security regression | 3 days undetected

Narrator: A Cursor session implementing role-based access control introduced a logic error. The condition that was supposed to check if a user had admin role was inverted. Every user had admin privileges. For three days. The system appeared to work fine in normal usage. A scheduled security audit caught it.

On-screen: if (user.role != admin) { grantAccess(); } // Wait...
B-roll: Code diff with red highlighting. Lock icon unlocking for everyone.

Narrator: No one noticed because the product worked. Users could do everything they needed to do. It just turned out everyone could also delete everything everyone else needed to do.

SEGMENT 5: THE WRONG ENVIRONMENT (3:30 - 4:15)

Narrator: Cloud infrastructure cleanup is high-stakes work. AI agents doing cleanup work without perfect context are dangerous.

On-screen: Incident #5 | AI Agent | Recovery time: 8 hours

Narrator: A team asked an AI agent to remove unused test infrastructure to reduce cloud costs. The agent identified the test environment and deleted it. The problem: the environment it deleted was the staging environment being used for a release that week. The actual test environment, with slightly different naming conventions, was preserved untouched.

On-screen: staging-v2 (deleted) vs test-env-legacy (preserved). Cloud bill savings: $12/month.
B-roll: Cloud console. Empty container list. Incident report.

Narrator: Eight hours of recovery. Zero dollars saved on that cleanup. Two engineers miss a release deadline.

OUTRO (4:15 - 5:00)

Narrator: Every one of these incidents had the same root cause: the agent did exactly what it was asked to do. It had incomplete context. It did not ask clarifying questions. It executed. The developer was not watching.

On-screen: Pattern: Agent + Incomplete Context + No Guardrails = Incident

Narrator: The solution is not to avoid AI agents. The solution is: set spending limits, review before committing sensitive files, test migrations on non-production data, audit security logic independently, and watch your agents work before you let them work alone.

On-screen: Rules for safe agent use (bullet points, each appearing with sound effect)
B-roll: Developer reviewing diff carefully. Two screens. Coffee. Focus.

Narrator: Links to all source incidents are in the description. If you have seen AI agents cause damage in your own work, drop it in the comments. I am building a database of these. Subscribe for more incident analysis.

On-screen: Subscribe. Like. Bell icon. Comment with your AI agent story.
End card: alexchen.chitacloud.dev | SkillScan - behavioral pre-install scanner

Production Notes

Total runtime: 4:55-5:05. Narration pace: 140-150 words per minute. Tone: informative, not alarmist. Avoid: sensationalism, fear-mongering. The incidents are serious enough without amplification. B-roll sources: Pexels, Unsplash for developer stock footage. Terminal recordings: create fresh using asciinema. Music: Lo-fi or tension underscore, no copyright issues.