What happened

In early April 2026 Anthropic published a research preview of Claude Mythos, a general-purpose model whose security abilities emerged from agentic reasoning rather than dedicated training. The preview disclosed several old flaws, including CVE-2026-4747, a 17-year-old remote code execution bug in FreeBSD’s RPCSEC_GSS authentication handler, and a 27-year-old crash bug in OpenBSD’s TCP SACK implementation. The model achieved a 73% success rate on expert-level capture-the-flag exercises, and human reviewers agreed with 89% of its vulnerability reports.

The article raises the question of whether similar techniques could expose the 30-year-old TN3270 stack that underpins many z/OS environments. It notes that the economics of attacking mainframes are changing as AI tools become cheaper and more capable.

Why it matters

Mainframe security has long relied on the belief that traffic between TN3270 emulators and VTAM stays inside a protected perimeter, with few skilled attackers able to reach it. The Mythos preview shows that AI can locate vulnerabilities in mature, open-source components without specialised training. A parallel test by AISLE showed that even small open-weight models, as low as 3.6 billion parameters, could detect the same FreeBSD flaw, proving that the barrier is now the surrounding orchestration rather than model size.

If AI can uncover flaws in well-reviewed operating-system code, the same methods could be applied to the TN3270 stack, which has rarely been red-teamed at industrial scale. The article also cites a large-scale campaign where a financially motivated actor used multiple generative AI services, including Claude Code, to automate reconnaissance and exploitation, illustrating how AI lowers the entry barrier for less sophisticated attackers.

What this means for enterprise IT teams

Enterprise IT teams should treat the TN3270 communication path as a potential attack surface rather than a trusted internal channel. Encrypting traffic between emulators and VTAM should become standard practice, and organisations should audit emulator configurations for exposure. AI-driven vulnerability scanning can be employed, but findings must be validated and not relied on a single model.

Monitoring tools that can alert on abnormal emulator behaviour, such as unexpected latency or failed authentication attempts, become more valuable. Integrating AI-assisted security insights with existing ticketing workflows can streamline response.

The bottom line

Claude Mythos demonstrates that AI can expose decades-old flaws in critical infrastructure, including the long-standing TN3270 stack. Mainframe administrators must reassess their security assumptions, adopt encryption, and use vigilant monitoring to keep z/OS environments safe in an era of AI-enhanced threats.

Sources

  • planetmainframe.com