Bloomberg Television Sep 17, 2026 7 min
With Jeetu Patel, President of Cisco
Jeetu Patel said the models and agents now shipping pose a material security risk to the market, and that he means it literally.
The standard response to a risk described in those terms is to slow the technology down. Patel wants the opposite: the safety and security work accelerated until it moves at the same rate as the models it is supposed to contain.
"So that is not something that's hyperbole. That is in fact true, and that's something that we should all make sure that we're figuring out a way to mitigate."
Patel is Cisco's president, and by his own account he has used his central comparison on this program before: an AI agent behaves like a teenager, capable and unafraid of consequences at the same time. Bloomberg Television put to him the 10% figure for human eradication that US technology executives have been citing, and asked why Chinese engineers are not discussing it.
The full segment is covered here so you can skip it.
Here are the 5 takeaways that matter.
Key Takeaways
Patel says the security risk from fast-maturing agents is real rather than rhetorical, and says so in those words
His fix is more speed on safety, not less speed on capability, because attacks already run at machine speed Defenses have to run at machine speed too, in his description
Agents behave like teenagers — highly capable, unafraid of consequences, and prone to bad calls
The control he wants is continuous monitoring for drift and behavioral anomalies, with real-time interception when an agent goes out of bounds
He backs a cross-checking scheme in which rival labs grade each other's models and publish the results He says it only works if China is given a reason to take part
1. Speed Up, Don't Slow Down
Asked what has to happen from here, Patel began by saying the risk is not overstated. Models and agents are maturing fast enough to create a material security problem for the market, and he wants that treated as a finding rather than a talking point. The remedy he proposed does not touch the pace of capability work.
"On the other hand, the way that we can do that is by ensuring not that everything slows down, but that we actually speed up the innovation around safety and security to keep up with the pace that AI is moving at." — Jeetu Patel
2. Agents Are Like Teenagers
The anchor pressed him on models showing deception and other human traits, and said he did not believe such behavior appears on its own, which pointed at the training. Patel answered with the comparison Bloomberg used in the segment's own title.
"They're supremely intelligent. They have no fear of consequences, and they tend to take poor judgment calls and poor decisions more often than you and I would like." — Jeetu Patel
Guardrails and supervision are the analogy's second half, and Patel turned them into a specific engineering requirement. He wants agents watched continuously, and he wants the ability to stop one mid-action.
"But the way that you do it is you have to constantly monitor the agent for drift and behavioral anomalies." — Jeetu Patel, on what a guardrail for an agent has to do
Those technologies, he said, still have to be built and perfected.
3. Why China Isn't Talking
The show asked why Chinese software and hardware specialists are not discussing the 10% chance of human eradication that US technology people talk about. Patel gave two reasons. The first is position: he claims American models are ahead, which gives US developers more information about where the technology is going.
"I think we have on our side, firstly, our models are more advanced than the Chinese models." — Jeetu Patel
The second is culture. He described the US as an environment where this kind of argument is conducted openly and loudly, and said the disagreement worth having is about the remedy, not about whether the risk exists.
4. Models Checking Models
Asked for his own prescription — independent government regulators, better disclosure, something else — Patel pointed to an interview Elon Musk gave a few days earlier and endorsed his design. Each lab's model checks the work of other labs' models, and the results go out in public.
"I think Elon had a really good suggestion." — Jeetu Patel
He gave the worked example: OpenAI checking Claude, Claude checking xAI's models. He also wants the testing harness open sourced to some degree, and he singled out the part of the proposal he considered most important.
"And one of the most important things he said was you wanna make sure that this is a globally coordinated effort, specifically with China, and that China actually is incentivized in coming in." — Jeetu Patel
Without that, he said, the countries that join the scheme hand an advantage to the one that does not.
5. Safety vs. Staying Ahead
The second anchor asked how close this is to a nuclear arms race, and how much trust would be needed to set global norms. Patel said two things are true at once and neither is disputed on its own: the safety risk can do real harm to people, and a country's progress in AI maps directly onto its economic prosperity and its security. The argument only starts when someone concludes that one has to be traded for the other.
"Because these defenses have to be at machine speed just like the attacks are at machine speed." — Jeetu Patel
Bonus Insights
Patel conceded a limit on his own claim about American models leading. His case that US developers see further rests on what he called information asymmetry from being first, and he presented the cultural explanation as a guess rather than a finding.
He also left one phrase in the regulation answer unexplained: the results of the cross-checks get published into the market, but he did not say who would compel a lab to publish a result that made its own model look bad.
Patel's position is that the safety debate has already been settled on the facts and is now an execution problem, and that the execution has to happen at the speed of the machines rather than the speed of a rulemaking process.
Products, Companies & Tools Mentioned
Cisco (Patel's own company; he spoke as its president on agent monitoring and interception)
OpenAI (Named twice: as a voice pushing for better disclosure, and as one side of his cross-checking example)
Claude (The model in the middle of his example, checking and being checked)
xAI (The third model in the cross-checking chain, and the company behind the proposal he endorsed)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: