Dr Heidy Khlaaf posted a thread stating that use of AI for cyber defense introduces unmitigated attack vectors. The thread includes a generated headline about prompt injection tricking Claude code into running malicious binaries. a16z quoted Greg Brockman describing how OpenAI directed 25 percent of production engineers to use models for finding vulnerabilities until none remained. Gary Marcus warned that defensive AI is not magic and pointed to the thread. David Sacks argued the solution to AI cyberattacks is more AI-powered defense, while other posters debated model capability limits and security dynamics.