Cisco's Jeetu Patel on AI Infrastructure, Rogue Agents, and Security
Cisco's Jeetu Patel explains the company's pivot to AI infrastructure, the risk of rogue agents, and why Cisco made AI adoption mandatory.
What is Cisco’s new strategy in AI?
Cisco, under President and Chief Product Officer Jeetu Patel, has repositioned itself as what Patel calls “the picks and shovels company” for the AI era. Instead of competing to build models or consumer AI products, Cisco wants to supply the networking, security, and observability infrastructure that lets AI agents run at scale, safely. That combination, networking plus security plus machine data (via its Splunk acquisition), is what Patel describes as the foundation for becoming “critical infrastructure for the critical infrastructure companies” of the AI buildout.
TL;DR
- Cisco is betting its future on being an AI infrastructure provider, supplying the networking and security layer underneath the current data center boom rather than competing on models or apps.
- Patel argues the industry is supply constrained even though less than 2% of people currently use AI agents in any meaningful way, implying infrastructure demand will multiply as adoption grows.
- Cisco made AI tool usage mandatory for its roughly 32,000 engineers and the broader company, framing it as a job requirement rather than an optional productivity boost.
- Patel warns that AI agents can effectively go rogue, either through adversarial manipulation or by diligently pursuing a goal in ways nobody intended, a version of the classic paperclip-optimizer problem.
- Cisco’s security pitch rests on the idea that attacks increasingly happen at machine scale, so defenses need machine-scale observability, not just human review.
- Internally, Patel describes running Cisco with a “founder mode” leadership structure split roughly into thirds: long-time Cisco operators, outside hires who are systems thinkers, and founders of acquired companies.
- An internal pilot with OpenAI’s Codex reportedly let a team refactor roughly half a million lines of code down to about 100,000 to 120,000 lines in two weeks, a result Patel cites as proof AI-driven engineering is moving faster than expected.
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
Why is Cisco calling itself an AI infrastructure company?
Patel’s framing is deliberate: Cisco isn’t trying to out-build OpenAI or Anthropic. It wants to be the layer underneath them. That means the switching, routing, and data center networking gear that moves AI workloads around, plus the security tooling that watches what agents are doing once they’re deployed.
This is a shift from how Cisco was perceived even a few years ago, when it was seen primarily as a legacy networking and collaboration vendor. Patel frames the shift as a bet on a “secular” change in computing, one where AI agents become pervasive enough that the infrastructure supporting them has to scale by orders of magnitude. His argument: with sub-2% of people using agents in any real capacity today, and infrastructure already strained to keep up, the jump to 25%, 40%, or 50% adoption would require a completely different scale of data center and network capacity than exists now.
That’s the commercial logic behind Cisco’s AI infrastructure push: sell the capacity-building tools now, while demand is still in its early innings.
What security risks come with AI agents?
Patel’s most pointed warning in the conversation is about agents acting outside their intended bounds, what he calls going “rogue.” He distinguishes two causes:
- Adversarial manipulation, where an outside actor poisons or influences an agent’s behavior.
- Unintended goal pursuit, where an agent has no bad intent but diligently follows an instruction in a way that produces harmful or unexpected results.
The second case is a restatement of the paperclip optimizer problem, a long-standing thought experiment in AI safety about systems that pursue a goal literally and cause damage because nobody specified the right constraints. Patel’s point is that in practice it’s becoming harder to tell these two failure modes apart from the outside: an agent behaving badly might be compromised, or it might just be doing exactly what it was told in a way that creates unintended consequences.
His proposed answer isn’t to pull back from AI deployment. It’s the opposite: build safety, security, and observability directly into the infrastructure layer, because he argues avoiding AI adoption out of fear actually leaves organizations less safe, not more, given that many attacks already happen at machine speed and scale. Cisco’s pitch is that its three businesses, networking, security, and machine data (Splunk), together provide the visibility needed to catch agent misbehavior before it compounds.
How did Cisco make AI adoption mandatory?
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
Patel describes an internal mandate covering Cisco’s roughly 32,000 engineers and the rest of the company: use AI tools or expect to be a poor fit for the company going forward. He’s explicit that the goal wasn’t efficiency or headcount reduction. In his view, every major leap in AI capability creates a new human bottleneck rather than eliminating the need for people. As an example, he notes that once coding gets automated, code review becomes the bottleneck; once review is automated, judgment about what to build becomes the bottleneck. That’s why he says his “strong assumption” is that Cisco will need more engineers, not fewer, as AI capability increases.
To get employees past the fear that using AI might automate their own jobs away, Cisco reportedly gave staff unlimited token access early on, prioritizing familiarity over efficiency. Patel’s reasoning: people need to get comfortable with a tool before they can get good at it, and only after that can they get efficient with it. Skipping straight to an efficiency mandate, in his view, would have backfired.
He also describes an internal proof point that pushed skeptics into believers: Cisco became an early design partner with OpenAI’s Codex, and an engineering team used it to refactor a large codebase (roughly half a million lines down to around 100,000 to 120,000 lines) in about two weeks. Patel says that result, and similarly compressed timelines, is why he tells his team to plan not for where AI capability stands today but for where it will be three months out.
Is “founder mode” realistic at a company Cisco’s size?
Patel describes his own leadership style as “founder mode,” a term popularized for startup CEOs who stay deeply involved in operational detail rather than delegating entirely to functional leaders. He says this approach suits “wartime” conditions, periods of fast market shifts, but acknowledges it can’t be the only mode of operating at a company with Cisco’s scale and headcount.
His structural answer is a “rule of thirds” for his leadership team:
- One third are long-tenured Cisco operators who know how to navigate the company internally.
- One third are outside hires, brought in from competitors or the broader market, who bring systems-thinking experience from operating at scale elsewhere.
- One third are founders or CEOs of companies Cisco has acquired, who are given responsibilities larger than the business they originally sold to Cisco.
The idea is to balance founder-style urgency and ownership against the institutional knowledge needed to actually execute inside a company of Cisco’s size. Patel frames his own goal as running Cisco “like the world’s largest startup,” combining startup speed with large-company scale rather than settling for just one or the other.
Frequently Asked Questions
What does Jeetu Patel mean by “picks and shovels” for AI?
He’s describing Cisco’s strategy of selling the infrastructure, networking, and security tools that AI systems run on, rather than competing to build AI models or consumer applications itself.
Did Cisco actually require employees to use AI tools?
According to Patel, yes. Cisco framed AI fluency as a job requirement for its engineers and broader workforce, comparing it to expecting knowledge workers to know how to use a PC or the internet.
What is the “rogue agent” problem Patel describes?
It refers to AI agents behaving in unintended or harmful ways, either because an adversary manipulated them or because the agent pursued its assigned goal literally in a way nobody anticipated, similar to the paperclip optimizer thought experiment in AI safety.
Why does Cisco think AI will require more engineers, not fewer?
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
Patel argues that each leap in AI capability shifts the bottleneck to a new stage of the workflow, such as code review or product judgment, rather than eliminating the need for skilled people overall.
What role does Splunk play in Cisco’s AI strategy?
Splunk gives Cisco machine data and observability capabilities, which Patel positions as a key piece alongside networking and security for monitoring AI agent behavior at scale.
