Palo Alto Networks CEO: "AI Found 5 Years of Bugs in 6 Weeks"
In a Nutshell
Palo Alto Networks CEO Nikesh Arora claims AI tools like Mythos uncovered 5-7 years of code vulnerabilities in just six weeks on their codebase, showing AI will dramatically accelerate both attack and defense capabilities in cybersecurity. He predicts massive disruption across SaaS categories—analytical tools face obsolescence while infrastructure and workflow systems get reinvented—as enterprises need 10x more data collection to defend against AI-powered threats. Arora argues the real profit pools lie in applications and cybersecurity rather than raw models, with Palo Alto positioned to capture margin expansion by using AI to run a more efficient enterprise.
These notes were generated by AI and may contain inaccuracies.
Palo Alto Networks is an outperformer in the cybersecurity space. CEO Nikesh Arora has been in the role for nearly eight years. When he started, the market cap was $17 billion. As of the interview date, the market cap reached $238 billion. The first 10x growth was the hardest phase; the company is now positioned for potential further exponential growth.
Nikesh Arora draws an analogy between AI and Google Search. Google democratized information; AI is democratizing intelligence. In marketing, 250 people produce varied outputs—AI can make 90% of that output consistent. With 5,000 customer-facing employees, the previous failure mode was inconsistency in problem-solving quality. AI enables near-consistent interactions across all representatives.
Mythos demonstrated that AI can assess all bad code written over the last 50 years and reveal vulnerabilities. In a six-week test, Mythos found vulnerabilities that would normally have taken 5 to 7 years to discover. The testing was performed on Palo Alto Networks' own codebase. When run in "ultra mode" with persistent thinking, Mythos can daisy-chain vulnerabilities to identify new attack paths. Palo Alto Networks already belongs to the top percentile of companies in code testing due to its cybersecurity focus. Across the broader industry and its 10 million developers, similar AI capabilities could surface issues that would have taken 10 years to find manually.
The cost of the Mythos testing run was in the low millions of dollars. Cost curves are declining rapidly, with newer models from OpenAI and Anthropic offering cheaper, more consistent performance.
IBM announced a $5 billion project to address open source vulnerabilities. If restraint on advanced models were removed, capabilities similar to Mythos could become available in the wild within three months. Chinese models and open models are expected to reach comparable levels around the same timeframe. An old industrial OT system running on the edge could be scanned for vulnerabilities relatively easily.
The industry is in a race between cyber defenders and attackers using the same AI tools. Enterprises are not performing as well as they should. CISOs face simultaneous pressure from vendors demanding patches for their products while also needing to address internal vulnerabilities. Open source remains a massive unsolved problem.
As model capabilities increase, systemic business risk for large enterprises also rises on the cyber side. Antidotes are being developed that may reduce the need to patch every vulnerability. Organizations currently lack memory and context of daily operations, requiring 10 times more enterprise data collection from a cybersecurity perspective to defend against AI-powered attackers.
Analytical SaaS companies face obsolescence. These are firms that collect data and sell analysis as a service. With direct model access, users can run their own analytics against stored data rather than purchasing incremental marketplace apps. One example involved a 20-seat SaaS product with low usage; three accounts were created, connected to Slack and Claude for natural language access, and the bill was reduced by 90%.
Infrastructure software (databases, data lakes, storage systems such as Databricks, Snowflake, MongoDB, Oracle) will see increased demand. Enterprises will need to store 10 times more data over the next three years.
Systems of work / systems of record (deeply embedded workflow software) will require re-engineering. UI layers will diminish as agents handle tasks. Sales representatives could dictate call outcomes to agents that complete all downstream paperwork. Audit trails improve because humans no longer touch data directly. The entire category is expected to be reinvented over the next five years.
89% of breaches still occur due to stolen credentials or simple password issues rather than sophisticated model-driven attacks. Models will eventually be capable of attacking critical infrastructure. National security agencies are considered well-resourced for their own protection. The greater concern is small offices and service providers (dentist offices, physician practices) using packaged software. The Change Healthcare breach caused ransomware-driven shutdowns across physician offices and required UnitedHealth to issue billions in credits.
The primary risk is economic chaos rather than targeted attacks on major infrastructure like PG&E power generation.
Nikesh Arora believes models will evolve into a utility layer with differentiated intelligence available on demand at varying price points (e.g., 120 IQ vs. 250 IQ tasks). Profit pools reside in applications, not raw model usage. Most companies lack the expertise to use models directly. OpenAI and Anthropic are viewed as the new Microsoft Office layer, but application companies will likely arbitrage between models to solve specific business problems. The application layer remains underdeveloped; 50,000 companies need the same applications rather than each building their own.
Entire frontier model weights can now fit on a USB stick, with data distillation possible in 24–48 hours. This makes long-term withholding of advanced models impractical in a global race.
False positive rates represent a critical undiscussed issue. Mythos exhibited a 30% false positive rate. Using raw models without proper harnesses and training can produce 10–20% false positives. In insurance claims processing, such rates translate directly to financial losses. In cybersecurity, the target is 0% false positives without increasing false negatives. The analogy is drawn to self-driving cars: a 10% false positive rate would be unacceptable for passenger safety.
Significant post-model work (harnesses, memory, context, training) is required to make models enterprise-ready.
Anthropic has improved ARR much faster than OpenAI by going all in on enterprise and security specifically. The current conversation centers on a race to take over profit pools, with companies needing tens of billions of dollars annually. One gigawatt of capacity generates approximately 10 billion in revenue.
Coding has been the breakout application over the past year and is massive. Infrastructure includes new databases. Cybersecurity is clearly a major profit pool because threats and patching cycles are far more dynamic than before.
AI models are positioned as enablers of better cybersecurity, which benefits everyone. Anthropic has already made its cybersecurity-capable model available generally so everyone can use it. OpenAI has one as well, and Google is expected to follow. CISOs and chief security officers want to use these models to test code.
Application software represents tens of billions of dollars waiting to be reinvented. Companies are exploring whether taking a $40-50-100 billion TAM and building a completely new AI backbone could create differentiation sufficient to move customers.
Year zero and year one companies are pitching $1,000 per seat per year or $500 per month SAS software that they can deliver at lower cost through consumption-based pricing, removing 80-90% of the cost. The two fastest places to generate revenue in enterprise are replacement TAMs, where an existing budget already exists and replacing something bad with something better is straightforward. The second area is consumer revenue, where acquiring $5 per user is relatively easy.
Hardware remains the cheapest way to manage low latency and high throughput bits. Data centers fundamentally manage high throughput and low latency bits. Financial services is the most reluctant industry to move to the cloud because increased latency directly reduces profit. Major financial institutions including Goldman Sachs, JP Morgan, and Morgan Stanley continue running hardware because cloud latency would cause them to lose money.
Hardware companies remain viable. Dell was once considered finished but has returned to approximately a $300-400 billion market cap. Hardware development cycles are being impacted by AI, with generative design moving into silicon that was historically manual and long-cycle.
The long pole in the tent for hardware is not design but production. Hardware componentry is backordered, expensive, and factories worldwide are backordered due to demand for GPUs and chip cards for data centers globally.
With firm top-down commitment, the US could address supply chain needs within 10 years. The hardware industry is experiencing a bonanza of a lifetime, enabling companies to commit $10, $20, $50, or $100 billion. One CEO publicly committed to a $100 billion plan to build more memory. Tax incentives including 100% accelerated depreciation write-off in the first year are driving these investments.
Over the past eight years, Palo Alto Networks has grown both organically and through acquisitions. Until approximately 18 months ago, the company bought product companies and integrated them into its go-to-market engine, rewiring their backend systems. This enabled selling $10 million to a customer and then $20 million two years later, representing the most efficient use of go-to-market spend. This playbook supported growth to $15 billion.
The company then identified an inflection in identity that would be critical from both agentic and security perspectives, leading to the acquisition of a $25 billion company that closed three months prior.
A new opportunity has emerged: being the best at leveraging AI to run the most efficient enterprise business can produce operating margins far exceeding industry averages. Achieving gross margins in the 90s and net margins in the 40-50% range would make the specific acquisition target less critical. Most subscale companies cannot afford to optimize their operations. If Palo Alto Networks can run with higher operating margins than competitors, the market will accept acquisitions of businesses at 20% margins and improve them.
The first M&A was met with skepticism because the acquirer lacked cybersecurity experience coming from Google, where track records of alumni success are mixed. The menu of potential acquisitions is now open, but the next 6-12 months are needed to understand how AI settles and how it can be used effectively in enterprises.
The view that AI will reduce headcount is countered by the belief that Palo Alto Networks will have more people on the technology side than ever before. AI is driving transformation across everything, requiring more technical personnel than would have been needed without AI.
Keep All-In Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
