Menu

1 / 2

Cisco Merged My PR in 39 Minutes — Why Prompt Defense Is the Next SQL Injection

DEV Community·ppcvote·6 months ago
#g9mt4vyl
Reading 0:00
15s threshold

39 Minutes

That's how long it took Cisco AI Defense to go from receiving my PR to merging it into main.

An 873-star repo ( cisco-ai-defense/mcp-scanner ). 27 minutes to approval, 12 more to merge. I was on a subway watching GitHub notifications, hands shaking enough I almost missed my stop.

But this post isn't about those 39 minutes.

It's about the four months that made those 39 minutes possible.

The Trigger: A Casual Scan

Rewind to January 2026.

I was building UltraProbe — an AI security scanner. One core function: check whether LLM system prompts have basic prompt-injection defenses.

I thought: "Let me dogfood this. Run it across a hundred or two public prompts."

After the scan completed, I stared at the screen for five minutes.

78% scored F.

Not "could be designed better" F. No defensive language at all F. No role-escape mitigation, no output-manipulation guards, no input-validation boundaries. Nothing.

Including some prompts I'd written myself a few weeks earlier.

It was a strange moment.…

Continue reading — create a free account

Join HashtagPLUS to read full articles, follow hashtags, vote, and join the conversation.