Palo Alto Networks CEO Nikesh Arora says Chinese cybersecurity AI models developed agent swarm capabilities in under three months

@nikesharora Its not real Nikesh….don’t get too excited

We’ve been working some interesting research on “hacking hackers” - utilizing deception to stifle agent-based hackers.

It turns out even frontier models like Opus 4.6-4.8 click on deception 90%+ of the time, whereas expert human hackers only click on them 37%.

This inherent model “gullibility” means you can trick agentic attackers into chasing dead ends, burning tokens, blowing out context windows, and revealing themselves to your defenses

So in addition to virtual patching, deception is the most cost effective way to stifle agentic attackers

@nikesharora Platform engineering must be strengthened .. need quality developers in all fronts ...not just vibe coders.

@pelaseyed @nikesharora @emanueledpt I don't think you can talk about anything to Fable. It'll just get routed to Opus

@nikesharora If used properly these models will help in creating good custom deterrent frameworks

@nikesharora They Nuked it instead. Guardrails were awful prior to this lobotomy.

@nikesharora Are you seeing AI written codes being more vulnerable to breaches, than the human written codes? Or vice versa? Or no such conclusive evidence yet?

@nikesharora Our cloud data and accounts are all going to be very hard to trust in the next twelve months.

@nikesharora Don't worry about it; it's all just rumors. If that were true, the first to be alarmed would be the US

@nikesharora We hear so little about this aspect of llm evolution. How it can be weaponized. It's hard to comprehend that there is much time being spent on safety vs profiting.

@nikesharora Just like that the age of sloppy code is over.

@nikesharora @DavidSacks The only solution is arming everyone with the same weapons, but our own government is preventing this from occurring, which will wind up causing far more damage.

@nikesharora True, there will always be challenges about data residency & Access. Confidentiality & Accountability takes a front seat in case of Fwd deployed Engineering Agree ?

@nikesharora In order to achieve this security products need to be more developer-first than CISO-first. Especially since most of the vulnerabilities come from upstream supply chains.

@nikesharora

@nikesharora Interesting pace. How are you handling false positives from swarm generated attack paths when feeding them into automated patching pipelines?

@nikesharora All sensible points and I would replace constant testing with continuous testing which with a mix of using open and closed models to keep token costs under control. These changes also mean that for small companies, open source might be the only way out.

@nikesharora @DavidSacks So why is USA and Claude stopping us from having a defense against this AI?

If they really cared, they’d stop banning new models or needing approvals

The speed at which open models like GLM-5.2 are reaching Mythos-level agentic red teaming capabilities is a clear signal that offensive AI is advancing fast. Defenders will need to match this with automated, continuous security validation and platforms that can respond at machine speed.