Causal Security's Cipher agent finds four critical zero-days in particle-accelerator control software in under 4 hours

The agent and the find

The Register reported on 28 October 2026 that CausaL Security's autonomous AI agent Cipher independently discovered four zero-day vulnerabilities in EPICS Base release 7.0.10 and earlier — the open-source control toolkit that runs roughly 90 particle accelerators and fusion reactors worldwide, including Argonne National Laboratory's Advanced Photon Source (APS) and the ITER fusion experiment. The agent completed the task in 3 hours and 49 minutes, requiring one authorization for deeper testing but zero initial human intervention.According to co-founder Shubham Raj, Cipher's approach is fundamentally different from traditional fuzzing or static analysis: the agent reason about where bugs might be in the codebase, crafts exploit payloads to verify each candidate, and iterates autonomously. For the EPICS Base target, it identified four distinct vulnerabilities across the control software's protocol handling and memory management.

The vulnerabilities

The critical finding is CVE-2026-114595, a CVSS 9.8 integer-overflow vulnerability that allows remote code execution. An attacker exploiting it could take full control of the EPICS control system, potentially redirecting accelerator operations or shutting down critical experiments. Three additional high-severity bugs were disclosed: one rated CVSS 8.2 and two rated CVSS 7.5.

CISA published advisory ICSA-26-300-01 on 27 October 2026 covering all four vulnerabilities. The agency says users should update to EPICS Base 7.0.11 and reports no known public exploitation as of the advisory date.

What it means

The Cipher result is one of the clearest demonstrations yet that autonomous AI security agents can find previously unknown critical vulnerabilities in real, complex software — and do so faster than typical human-led discovery cycles. Four zero-days in under four hours, including a critical RCE, is a pace that redefines what is possible in vulnerability research.

For the scientific infrastructure community, the immediate takeaway is operational: EPICS sites should update to 7.0.11. For the security industry, the signal pushes in a harder direction: adversaries with access to similar agents can now discover exploits at this pace too, compressing the window between disclosure and exploitation.

The broader context

CausaL Security has been developing Cipher as a platform for autonomous vulnerability discovery, and this EPICS disclosure is its highest-profile result to date. The agent's ability to reason about control-system software — a domain with unusual protocol code, constrained memory models and limited fuzzing infrastructure — demonstrates general capability, not narrow overfitting to web applications.

The finding also raises the question of disclosure norms for agent-found vulnerabilities: CISA coordinated this disclosure through the standard ICS advisory process, but the volume of agent discoveries is likely to grow, and the treatment of automated security research within vulnerability disclosure programs remains uneven.

Caveats and what to watch

The 'autonomous agent' claim should be read with the qualification that one authorization was needed during testing, which falls short of fully unattended operation. The discovery time and vulnerability counts are reported by the vendor, and independent replication has not been performed.

Watch whether similar agent-driven zero-day discoveries accelerate across ICS, embedded and enterprise software in the coming months. Watch for disclosure policy responses from CISA and other agencies as agent-driven discovery volumes increase.

How Cipher works

CausaL Security has been developing Cipher as a platform that uses large language models to reason about source code, identify potential vulnerability patterns and generate and refine exploit payloads. Unlike fuzzers that need seed inputs and mutation strategies, or static analyzers that produce alerts without confirmation, Cipher treats the vulnerability discovery process as an autonomous reasoning loop: read code, hypothesize a bug location, craft an input to test it, observe the result, iterate.

The EPICS Base result is significant partly because critical-infrastructure code is notoriously difficult to test with conventional tools: control-system protocols have unusual state machines, memory models that differ from web application code and limited unit-test infrastructure. Cipher's ability to work in that domain suggests the approach generalizes beyond web-app security.

Implications for the security industry

The most direct implication is that the economics of zero-day discovery have changed. A human security researcher or team might spend weeks or months finding a critical RCE in industrial control software. An autonomous agent that finds four in under four hours, even allowing for vendor claims and the need for one human authorization, compresses the discovery cycle by orders of magnitude.

For defenses, this means the window between a vulnerability being discoverable and it being exploited narrows further. Vendor patch cadence and CISA advisory processes that assumed weeks or months of discovery-to-disclosure time will face pressure. It also means bug bounty programs and VDPs will need to decide how to handle agent-submitted findings — a question that is already live for several major platforms.

Caveats

The autonomous-agency claim is qualified: Cipher required one authorization during testing to achieve deeper access, which means the 'zero initial human intervention' framing is not fully unattended. The vulnerability counts and discovery time are reported by CausaL Security and have not been independently replicated.

The EPICS Base fix to 7.0.11 addresses the disclosed four CVEs; sites should update, but the discovery does not imply additional undisclosed vulnerabilities at the same severity level.