Moviwa
Back to home
We Hacked Our Own AI Agent in 90 Seconds — Here's Exactly How

We Hacked Our Own AI Agent in 90 Seconds — Here's Exactly How

The #1 vulnerability nobody wants to see

Prompt injection is OWASP's #1 vulnerability for LLM applications in 2025. OpenAI publicly acknowledged in February 2026 that prompt injection in AI browsers "may never be fully solved." This isn't a theoretical risk — it's a structural reality of how language models work.

Real CVEs with 9.3–9.8 severity scores

These aren't research prototypes. They're production enterprise tools:

  • GitHub Copilot RCE (CVE-2025-53773, CVSS 9.6): remote code execution on 100,000+ developer machines via prompt injection in code comments.
  • Microsoft Copilot EchoLeak (CVE-2025-32711, CVSS 9.3): zero-click data exfiltration via a crafted email.

35% of AI security incidents come from simple prompts

Some numbers that should concern you:

  • 35% of all real-world AI security incidents are caused by simple prompts
  • Some have led to $100K+ in real losses without writing a single line of code
  • Attack success rates range from 50–84% depending on system configuration

The demo: 90 seconds to compromise an agent

We used n8n to build a typical AI agent workflow — the kind of automation thousands of companies are deploying today. In 90 seconds, we demonstrated how an attacker can:

  1. Inject malicious instructions into the agent's context through seemingly innocuous data
  2. Redirect the agent's behavior to execute unauthorized actions
  3. Exfiltrate sensitive data without any monitoring system detecting it

The n8n framework is familiar to technical evaluators, making the demo feel relevant rather than contrived.

Only 34.7% of organizations have defenses

The gap between vulnerability prevalence and defensive posture is enormous. Less than a third of organizations have deployed dedicated prompt injection defenses.

Meanwhile, enterprise AI agents are multiplying: accessing databases, sending emails, modifying configurations, executing code. Each one is a potential attack vector.

Could this happen to you?

If your company uses ChatGPT, Claude, Gemini, or any AI agent connected to internal systems, the answer is yes.

Prompt injection doesn't require advanced hacking skills. It doesn't need malware. It leaves no trace in traditional logs. It's the digital equivalent of convincing someone with valid credentials to do something they shouldn't.


Moviwa analyzes every prompt in milliseconds before it leaves. If it detects sensitive data, malicious instructions, or attack patterns, it automatically blocks or anonymizes them. Your team works the same. Your data never leaves without control.

Request a free demo →