AI 'hacker' finds 4 severe bugs in popular industrial control software in under 4 hours

An autonomous AI agent discovered four security holes in the widely used EPICS open source toolkit, including one critical flaw.

This critical vulnerability, rated 9.8 under CVSS v3.1, is an integer overflow issue (CVE-2026-114595). It permits an attacker to achieve an out-of-bounds heap write, potentially leading to remote code execution.

The toolkit, which supports major scientific endeavors, is affected by four issues. These include the critical overflow, a high-severity message-length validation flaw (CVE-2026-114596) causing memory writes and crashes, a high-severity bounds checking bug (CVE-2026-114597) allowing memory reads and crashes, and a high-severity type confusion flaw (CVE-2026-114598) that risks denial of service and memory exposure.

CISA issued a Tuesday advisory, warning that "Successful exploitation of these vulnerabilities could allow an attacker to cause a denial-of-service condition, corrupt memory, disclose sensitive information, or potentially execute arbitrary code."

Causal Security co-founder Shubham Raj reported that his AI agent, Cipher, found all four previously unknown security holes in just three hours and 49 minutes. Raj stated, "The initial investigation required no human intervention, followed by one instruction authorizing deeper exploitability testing in a secure environment."

EPICS, originally developed by Los Alamos National Laboratory and Argonne National Laboratory, helps build and control SCADA systems globally. It is utilized in roughly 90 major scientific, university, and industrial projects, such as the Advanced Photon Source and ITER's fusion reactor control system.

Raj highlighted the danger of CVE-2026-114595, noting, "It allows attacker-controlled network data to cause an out-of-bounds heap memory write, with potential for remote code execution." He explained that because EPICS controllers manage commands and measurements between operators and physical gear, an attacker could potentially crash a controller during an experiment.

While reaching the vulnerable services on a control network is necessary, Raj noted that network misconfigurations could make this possible. He added that "Someone with conventional exploitation skills could investigate them, and increasingly capable AI systems could help develop exploits without requiring specialized OT expertise."