The next phase of AI cybersecurity still needs humans
TL;DR
Anthropic’s and OpenAI’s cyber-capable AI models still require significant human expertise to be effective, according to early users testing them in real environments. Anthropic’s Mythos Preview reportedly surfaced tens of thousands of bugs across nearly every operating system, with OpenAI’s GPT-5.5-Cyber matching it on bug discovery and exploit generation. The shift is less about fully autonomous hacking and more about how well humans can direct, validate, and operationalize these tools.
Nauti's Take
Good news for security teams: treating these models as accelerators rather than replacements brings real speed gains in bug triage and exploit research. The risk runs the other way — the same models will reach attackers, and without human validation they produce false positives and dangerous recommendations.
Teams with a clear review process benefit; anyone piping AI findings straight into pipelines should be cautious.
Sources