Build radar
Sources 33Updates always the freshest
0 Newsletter
Animated illustration for: An AI agent found four serious holes in lab software in under 4 hours
AI News · Oct 28, 2026 · 2 min

An AI agent found four serious holes in lab software in under 4 hours

An AI agent combed code that runs particle accelerators. I explain what it found, who is affected and what to do if you run similar software.

In short
  • Causal Security's AI agent found four previously unknown flaws in EPICS, software that controls scientific and industrial facilities.
  • The worst flaw scores 9.8 out of 10. The fix is in version 7.0.11.
  • CISA reports no known exploitation in attacks.

In short: on 27 October 2026 the US agency CISA published a warning about four flaws in EPICS, open-source software that controls scientific facilities. Shubham Raj of Causal Security reported them, and according to The Register they were found by an AI agent called Cipher.

What was found

  • Four flaws, all in EPICS Base version 7.0.10 and earlier.
  • The worst has a CVSS score of 9.8 (critical): a memory-size calculation error that can corrupt memory and potentially run foreign code.
  • The other three score 8.2, 7.5 and 7.5: mainly knocking a service offline and possibly reading neighbouring memory.
  • According to The Register, the agent found and confirmed all four in 3 hours and 49 minutes, with the first investigation done without a human. That timing comes from the company’s co-founder, not from an independent audit.

What to do

  • Update EPICS Base to version 7.0.11.
  • Keep control networks off the internet and behind a firewall that separates them from the office network.
  • Use a current VPN for remote access.

CISA says it knows of no public exploitation of these flaws. An attack requires access to a network where the vulnerable services run.

How to read this

The timing and the credit to the agent come from a company that sells such a tool, so I treat them as the vendor’s claim. The fix and the severity scores, however, come from the official advisory. I cover protecting code and agents in the AI code security hub.

Sources

AI code security

Related