09/24/2026
Two AI giants launched new models this week, and both say they resist prompt injection better than before.
The fine print tells a different story. OpenAI’s new model still followed planted instructions in about 1 in 9 test runs. Anthropic’s own report shows its new model got worse at one thing: obeying malicious instructions hidden in text you paste into your prompt.
Better isn’t fixed. Every release improves the odds, but models still read instructions and data as the same text.
If your plan is “wait for the next model,” you’re betting on a percentage. Test the AI you actually deploy.
DM me LAB for the hands-on prompt injection course: 5 labs, one afternoon, one payment.
aiagents owasp pentesting