5 AI Security Vulnerabilities

5 AI security vulnerabilities we found in production systems last month.

None of these companies thought they were exposed. All of them had "security" in place.

Here's what an audit actually found 🧵

Vulnerability #1: Prompt Injection

A customer support chatbot was instructed to "never reveal internal policies." A user pasted a fake system message that overrode the instruction.

Result: The bot leaked its entire system prompt, including internal project names and API key placeholders.

Fix: Input sanitization + separate instruction channels.

Vulnerability #2: Tool Access Misconfiguration

An AI agent had read-write access to a CRM. Through prompt injection, we convinced it to delete contact records — then summarize what it had done.

Result: 200+ records deleted in 4 seconds. No approval step existed.

Fix: Read-only permissions + human confirmation for destructive actions.

Vulnerability #3 & #4: Data Leakage + Model Extraction

A fine-tuned model was queryable without rate limiting. An attacker extracted training data by asking the model to "repeat what you learned in training."

Another team ran 10K queries to clone their model's behavior for $47.

Fix: Rate limits + output filtering + query auditing.

Vulnerability #5: Compliance Gaps

A startup passed SOC 2 but never tested their AI agent. The agent had access to PII with no audit trail. They would have failed if the auditor had looked.

Result: Remediation cost 3x more than an audit would have.

—

The pattern: Every company assumed their AI was "probably fine." None of them were.

If you've deployed an AI agent with tool access or customer data, you need an audit.

I wrote a complete guide on what's tested, how long it takes, and what the report looks like.

🔗 galeops.xyz/blog/ai-security-audit-guide

Or DM me "audit" and I'll tell you straight if you need one.