An AI Agent Shipped Four CVEs Into Industrial Control Software in Under Four Hours
CVE-2026-114595 is a CVSS 9.8 RCE in EPICS Base, and the finder is a product named Cipher
On Oct 27, The Register's Jessica Lyons reported that an autonomous AI agent called Cipher, built by Causal Security, found four severe vulnerabilities in EPICS Base — the open-source toolkit used to build and remotely control SCADA systems at roughly 90 scientific, university, and industrial facilities, including Argonne's Advanced Photon Source and ITER's fusion-reactor control system. Co-founder Shubham Raj told the publication the initial investigation took 3 hours and 49 minutes and required no human intervention on the first pass, plus one instruction to authorize deeper exploitability testing in a sandbox.
All four are now fixed in EPICS Base 7.0.11. CISA issued an advisory warning that exploitation could cause denial of service, memory corruption, information disclosure, or arbitrary code execution.
- CVE-2026-114595 — CVSS 9.8, Critical. Integer overflow driving an attacker-controlled out-of-bounds heap write, with potential remote code execution.
- CVE-2026-114596 — CVSS 8.2, High. Improper message-length validation, out-of-bounds memory write, controller crash.
- CVE-2026-114597 — CVSS 7.5, High. Insufficient bounds checking, out-of-bounds memory read, process crash.
- CVE-2026-114598 — CVSS 7.5, High. Type confusion, denial of service, possible adjacent-memory exposure.
Exploitation requires reaching the vulnerable EPICS service on a facility's control network — which happens when that network is misconfigured or poorly segmented, exactly the condition CISA has been warning operators about for years.
What's new is who the finder is
Autonomous AI vulnerability research is not a 2026 novelty. METR and Redwood Research found six ways to blind OpenAI's coding-agent monitor earlier this year. Anthropic's Oct 28 Frontier Red Team disclosure says an unreleased internal model chained flaws into five working cross-platform exploit chains against Chrome, Safari, Android, Windows, and Linux. On Sept 1, OpenAI announced it would limit release of its unreleased Astra model after classifying it as crossing the "Critical" cybersecurity-capability tier in its own Preparedness Framework.
Those are all frontier-lab self-reports. The Causal / Cipher story is a different shape:
- The finder is a named commercial agent product — Cipher, from Causal Security — not a research model behind a lab firewall.
- The target is named critical-infrastructure software, not a benchmark machine.
- The finding carries publicly-assigned CVE IDs with CVSS scores, not an internal research tag.
- There is a CISA advisory attached to it.
- The full research → disclosure → patch → public-advisory loop ran in a single reported episode, with the vendor confirming the fixes in EPICS Base 7.0.11.
Up to this run, the AI-agent-as-vulnerability-finder story in the public record was mostly vendor-controlled: Anthropic found a thing in Chrome, Anthropic reported it to Google, Google patched it, Google credited "Xinyang Ge of Anthropic, assisted by Claude." Now a non-frontier commercial agent is anchoring a CISA advisory on a widely deployed industrial control toolkit.
Why this changes the defender's math
The CVE curve bends. ThreatDown's Oct 28 "Patch Tsunami" analysis in THE Journal reports September 2026 saw 998 CVEs addressed around Patch Tuesday versus 96 a year earlier — roughly 10× year-over-year. The three-month rolling average of patched Linux kernel vulnerabilities tripled; Google Chrome's patched-vulnerability count rose 36×. The Verizon 2026 DBIR, cited in the same piece, says 35% of known actively exploited vulnerabilities were still unpatched after 28 days in 2025. That was before this.
The attacker's cost drops with the finder's cost. If a commercial agent can find a CVSS 9.8 RCE in an open-source industrial control toolkit in under four hours, so can a less friendly operator with a comparable tool chain. The research step that used to need a human expert and weeks is now priced near a cloud GPU hour.
The target set is wider than Chrome. EPICS Base is the kind of 20-year-old open-source library that sits deep inside systems nobody has fuzzed on a modern schedule. There are a lot of libraries like it.
What CISA's advisory doesn't do
The CISA advisory tells operators what the vulnerabilities are and how to patch them. It does not and cannot tell them whether anyone — or anyone's agent — is already running a Cipher-equivalent against their next upstream dependency. Facilities running EPICS on misconfigured control networks are on their own at that boundary.
The practical steps are the ones the EPICS maintainers and CISA already name:
- Upgrade EPICS Base to 7.0.11. All four CVEs are fixed there.
- Verify control-network segmentation. The exploit path for CVE-2026-114595 assumes the attacker reaches the EPICS service, which should not be from the enterprise or guest network.
- Treat anything new on the control-network wire that speaks the EPICS Channel Access protocol as something to notice. The specific primitives in CVE-2026-114595 and 114596 are at the message-parsing boundary.
- Expect more. CyberStrikeAI, tracked by TrendAI across 19+ organizations in Japan in October, has already demonstrated 44-agent kill-chain splits against enterprise targets. The industrial-control variant of that pattern is a matter of time.
The takeaway
A commercial AI agent shipped four CVEs — one at CVSS 9.8 — into a widely deployed industrial control toolkit in under four hours. CISA wrote an advisory. The vendor shipped the fix. Attackers now have access to the same shape of tool, at the same price. The question is which of the libraries sitting quietly inside your stack is next — and whether you hear about the patch before you hear about the exploit.
Watch the control network. Watch what the service process actually does when it parses a message it has never seen. The CVE curve is bending. The defender's observation surface has to bend with it.
SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.