Patrick Lucas Austin hosts IBM Technology’s 26-minute Security Intelligence panel with Austin Zeizel, Sophie Cunningham and Bryan Clark.

  • The panel discusses Claroc’s investigation of sensitive data exposed through Grok Build apps, including internal business documents and personal records.
  • Each app has its own hostname; passive DNS services, which collect domain lookups, can reveal addresses that users assumed were hard to discover.
  • Claroc reports that 35,578 of 37,571 reachable apps opened without sign-in; that count includes ordinary public apps.
  • Its test account proposed “Anyone with the link” at publication; the platform says employee apps default to private, so defaults need account-specific verification.
  • The panel recommends private access, approved internal sharing services and security review before sensitive data enters an AI-built app.

Building a tool also means maintaining its security · 5:43

  • Panelists speculate that speed-focused development can miss permissions; they do not establish that an AI chose the exposed settings in this incident.
  • A mature sharing service has staff responsible for access controls, while an improvised replacement can leave that work with an already-busy developer.
  • A panelist compares the problem to new carmakers rediscovering manufacturing problems that experienced companies have already encountered.
  • Generating a working interface leaves ongoing responsibility for identity, authorization and changes to the data it holds.

Plan an approved fallback before email fails · 10:05

  • The host attributes a Microsoft 365 outage to a gateway fix, hardware-dependent behavior, load and already-reduced capacity.
  • Zeizel says an outage can interrupt business similarly whether the cause is an attacker or an operational fault.
  • Cunningham warns that personal email, texts and ad hoc documents can move company data into unapproved services during an outage.
  • Continuity plans should name decision-makers and approved alternative channels before the primary provider fails.

Agents compressed a simulated industrial attack · 15:29

  • The panel discusses Shinobi Security’s reconstruction of a Polish wind-farm incident using a test rig and simulated turbines.
  • Researchers supplied VPN credentials; the agents exploited default credentials and configuration weaknesses rather than discovering new vulnerabilities.
  • The episode compares roughly two hours in the reconstruction with 11 days in the historical intrusion; these are different environments, not a controlled speed benchmark.
  • Computer Weekly’s report distinguishes Claude’s 66-minute run from Kimi’s just-under-two-hour run.
  • No live wind farm was shut down by this experiment; the reported result concerns the replicated environment.

Secure the boundary and rehearse the response · 17:52

  • Require multifactor authentication for remote access, replace default passwords and separate industrial control networks from other systems.
  • Prepare an operational-technology response plan, with operational technology meaning the systems that monitor and control physical equipment.
  • Panelists debate detailed public incident reports: they help defenders reproduce failures but can also lower attackers’ learning costs.
  • Coordinate disclosure and remediation; the model’s reported “safety theater” remark is an anecdote, not evidence of intent or general real-world autonomy.

“I don’t think it’s something you figure out when things go bad.” — Austin Zeizel, 17:57