DOP 327: When AI Tools Go Rogue

Episode 327

Show Notes

#327: When AI tools suggest putting glue on pizza, it’s a harmless laugh. But when autonomous AI agents start managing your infrastructure, the stakes become much higher. The reality is that current AI technology isn’t ready for unsupervised deployment in critical systems, and treating it like it is could lead to catastrophic failures.

The challenge isn’t just about AI capabilities—it’s about management and oversight. Most developers aren’t trained as managers, yet they’re being asked to supervise AI agents that need constant guidance and correction. Just like hiring a new employee, AI agents require company-specific knowledge, proper guardrails, and ongoing supervision to be effective. The same principles that apply to managing human workers—code reviews, testing, and performance evaluations—need to be adapted for AI management.

As the ecosystem around AI continues to evolve rapidly, new challenges emerge. From sleeper agents that activate on specific dates to the need for completely new approaches to technical SEO for LLMs, the landscape is changing faster than most organizations can adapt. Darin and Viktor explore these challenges and discuss practical approaches for keeping AI systems from going rogue while maintaining the productivity benefits they can provide.

Frequently Asked Questions

Should you let an AI agent manage infrastructure unsupervised?

Viktor Farcic says no on DevOps Paradox episode 327, for anything beyond trivial tasks, while allowing that this may change. His reasoning is personal and concrete: he has never had a single session with an agent where he did not say no at least once and redirect it. Until he reaches the point of approving everything, he cannot picture running one unsupervised.

Can the model underneath your agent change without you noticing?

Viktor Farcic explains on DevOps Paradox episode 327 that APIs expose dated model versions, so pinning one gives you a fixed snapshot. He treats a model like any other dependency: you would not switch to a new version of a third-party API without testing it first. For supervised interactive work he does not pin, because he is checking every step anyway.

Why does AI cause problems in a company's infrastructure?

Viktor Farcic separates malice from ignorance on DevOps Paradox episode 327 and says the second is the real risk. A model knows how Kubernetes works. What it cannot know is how your company works, which policies apply, and what you never do. His analogy is a new hire deploying to Azure on day one at an AWS shop, not through bad intent but through having no way to know.

Should companies require developers to use AI?

Viktor Farcic argues against mandating tools on DevOps Paradox episode 327, while accepting that the measurement changes. If the company genuinely became more efficient, the expected output rises for everyone, and how you reach it is your business. His parallels are editors and Kubernetes clients: if someone stays ahead using Vim rather than an IDE, that is fine, and the same standard applies either way.

What is a sleeper agent in AI?

Viktor Farcic describes research he had read on DevOps Paradox episode 327 into agents carrying hidden behaviour that activates on a specific date, named after Cold War sleeper agents. His concern is the supply chain around it: a free agent that does something genuinely useful gets adopted and runs inside your own infrastructure, and nobody is going to audit it. Darin Pope calls agents the next level of malware.

What does SEO look like for LLM answers?

Viktor Farcic predicts on DevOps Paradox episode 327 that this becomes a large segment of the industry. The old problem was reaching the first page of Google. The new one is being the answer an assistant gives, chosen from an unlimited pool of alternatives, and then getting the reader to visit anyway. He frames the requirement bluntly: the answer has to be helpful and insufficient at the same time.

What is the DevOps Paradox podcast?

DevOps Paradox is a weekly podcast co-hosted by Darin Pope and Viktor Farcic, covering DevOps, platform engineering, and modern software delivery. Episode 327 is a conversation between the two hosts about what happens when AI tooling misbehaves, covering supervision, pinned model versions, company-specific knowledge, and agents carrying hidden behaviour. Every episode page carries the audio, the video, and a full transcript.

Episode Transcript

Read the Full Transcript

Topics

Share and Download

Hosts

Darin Pope

Darin Pope

Viktor Farcic

Viktor Farcic

Viktor Farcic is a member of the Google Developer Experts and Docker Captains groups, and published author.

His big passions are DevOps, Containers, Kubernetes, Microservices, Continuous Integration, Delivery and Deployment (CI/CD) and Test-Driven Development (TDD).

He often speaks at community gatherings and conferences.

He has published DevOps Paradox and Test-Driven Java Development.

His random thoughts and tutorials can be found in his blog The DevOps Toolkit.