Palo Alto Networks CEO: "AI Found 5 Years of Bugs in 6 Weeks"

2026-06-08 - 31 min - source - Read full transcript

Nikesh AroraChamath Palihapitiya (host)Jason Calacanis (host)David Friedberg (host)

Key insights

Palo Alto's AI tool Mythos found 5-7 years' worth of code vulnerabilities in 6 weeks, for a low-millions-dollar token cost.

Arora says the test ran against decades of accumulated code and surfaced vulnerabilities at a pace no human team could match. He frames this as proof the capability is real, not hype, and notes the cost curve is already falling as newer models ship cheaper and more consistent.

ai-cybersecurity

Mythos-level vulnerability-finding capability could reach open/adversarial models within about 3 months, not the 6 previously assumed.

Arora points to already-released models (4.8, 5.5) with similar capabilities and argues attackers don't need to crack the hardest code, just find easy vulnerabilities in something like edge-running industrial OT code.

ai-cybersecurity

High false-positive rates, not raw capability, are the real blocker to using AI for cyber defense and other high-stakes enterprise tasks.

Arora cites an early model with roughly a 30% false positive rate: it is great for finding attack paths but unreliable for defense, since a third of flagged 'vulnerabilities' aren't real. He extends the point to any automated decision task (he uses insurance claims as an example) and to self-driving cars, arguing the post-training work to push false positives toward zero is where most of the real engineering effort has to go.

ai-cybersecurity

89% of breaches happen because of stolen credentials, not sophisticated exploits, and the biggest national-security exposure is small under-resourced operators, not hardened critical infrastructure.

Arora says agencies protecting critical infrastructure already spend roughly 10% of budget on security and are well staffed. The real risk sits with small offices, like a dentist or doctor's practice running packaged software, illustrated by the Change Healthcare ransomware breach that shut down physician offices nationwide and forced UnitedHealth to issue billions in credits.

ai-cybersecurity

Analytical SaaS is dead: once an enterprise's data is unified, an LLM can just be pointed at it instead of paying for a dedicated analytics app.

Arora defines analytical SaaS as any product whose value is 'collect your data and analyze it for you.' He argues that once a company has centralized data, it doesn't need a separate marketplace app to run analysis against it; it can run a model directly.

saas-disruption

Palo Alto killed a SaaS tool that only 3 of 20 seats were using, connected the underlying data to Slack and Claude, and cut that spend by 90%.

The tool had users unaware they even had data value; after collapsing licenses to 3 seats and wiring the data into a chat interface, the company got the same or better utility for a fraction of the cost. Arora frames this as the coming pattern across enterprise software.

saas-disruption

System-of-record software (Salesforce, Oracle, and similar systems of work) will get re-engineered over the next 5 years as UI disappears and agents interface with data directly.

Arora argues UI was 'the worst thing we did as technologists' because it exists only so humans can interact with data. If agents can capture a sales call, extract key points, and post them into the backend system without a human touching the UI, the audit trail improves and companies can rewire how work happens, turning five-person workflows into one.

enterprise-ai-transformation

Profit pools sit in the application layer, not in the underlying models, so an ecosystem of arbitrage applications will form between models and enterprise problems.

Arora points out that OpenAI's Codex, not OpenAI itself, is described as 'running away' with growth, and Anthropic's Claude Code gets the same framing. He argues models are becoming a fungible utility ('buy intelligence on the fly,' choosing IQ level and price per task) while the durable value accrues to the application companies that package models for specific business problems, a layer he says hasn't fully formed yet.

ai-profit-pools

Model weights are now small and portable enough (fitting on a USB stick) that trying to withhold a frontier model for national advantage doesn't meaningfully slow anyone down.

Arora relays a conversation with a model-company CEO who said their newest model's entire weights fit on a USB stick, meaning the IP can be distilled and reproduced in 24-48 hours once accessed. He uses this to argue against holding back capable models for 3-6 months as a policy lever, since someone else releases comparable capability in the open regardless.

ai-cybersecurity

Palo Alto is adding technical headcount, not cutting it, because AI is forcing every part of the business into transformation simultaneously.

Arora pushes back on the common 'AI means fewer people needed' narrative for his own company: because AI is prompting a transformation project across marketing, sales, and engineering all at once, Palo Alto currently has more technical staff than it has ever had.

enterprise-ai-transformation

Hardware and data centers remain essential because latency-sensitive workloads, especially in financial services, can't move fully to the cloud.

Arora says firms like Goldman, JP Morgan, and Morgan Stanley stay on hardware because added cloud latency directly reduces trading profit. He adds that today's bottleneck is production capacity, not design, since every hardware component is backordered amid the buildout, citing Dell's recovery to a $300-400B market cap as evidence hardware is far from dead.

ai-profit-pools

Arora's next M&A thesis is AI-driven operating leverage: if Palo Alto can push its own margins to roughly 90s gross / 40s net, that efficiency, not sector fit, becomes the justification for broader acquisitions.

Historically the company bought product companies and ran them through its go-to-market engine, most recently a $25B identity-security acquisition closed 3 months before this conversation. Arora frames a new logic: most subscale companies can't afford to optimize themselves with AI, so if Palo Alto proves it can run any acquired business at a much higher operating margin than peers, the specific industry matters less.

ai-profit-pools

Companies

Palo Alto Networks - Arora's company; built Mythos, the AI tool that found 5-7 years of code vulnerabilities in 6 weeks; closed a $25B identity-security acquisition 3 months before this taping
Google - Arora spent 10 years there (former chief business officer); he argues Google is underrated and could be the first trillion-dollar company, citing its sales-force scale
OpenAI - Referenced for model releases (4.8, 5.5) and the framing that its profit pools are in applications like Codex, not the base models
Anthropic - Referenced for Claude, its cyber-capable model release, and its faster ARR growth versus OpenAI from going all-in on enterprise
IBM - Announced a $5B project to fix open-source software vulnerabilities
Salesforce - Cited as an example of a SaaS marketplace app category that AI can replace by querying the underlying data directly
SAP - Cited as a source of inventory data an enterprise would want unified with sales data for AI-run analytics
Databricks - Named as infrastructure software that Arora sees as undervalued relative to analytical SaaS
Snowflake - Named alongside Databricks as core data-storage infrastructure that will stay valuable
MongoDB - Named as core database infrastructure that will stay valuable
Oracle - Named as core infrastructure and as a system-of-record example agents will eventually interface with directly
Dell - Cited as an example of hardware demand cycling back after being written off, now near a $300-400B market cap
Silver Lake - Arora previously advised the firm when hardware, including Dell, was considered a dying category
Change Healthcare - Ransomware breach that shut down physician offices nationwide, used as the example of the real national-security risk (small, under-resourced operators, not hardened critical infrastructure)
UnitedHealth - Had to issue billions in credits to physicians after the Change Healthcare ransomware breach
Uber - Arora sits on Uber's board and declined to give an assessment, praising CEO Dara Khosrowshahi
Waymo - Praised for reliability; Arora says it should expand to more cities faster
Goldman Sachs - Cited among financial-services firms that avoid the cloud because of latency sensitivity
JP Morgan - Cited among financial-services firms that avoid the cloud because of latency sensitivity
Morgan Stanley - Cited among financial-services firms that avoid the cloud because of latency sensitivity
Netflix - Cited as an example of consumer subscription revenue stacking up faster than people realize

Techniques and frameworks

Mythos - Palo Alto Networks' AI system for assessing code vulnerabilities; in 'ultra mode' (persistent thinking) it can daisy-chain vulnerabilities into full attack paths
False positive / false negative rate framing - Arora's lens for why raw model capability isn't enough for enterprise use; an early model's ~30% false positive rate makes it unusable for defense despite being usable for attack
Analytical SaaS vs. infrastructure vs. system-of-record - Arora's three-way framework for how AI hits different software categories differently: analytical SaaS dies, infrastructure software is undervalued, system-of-record software gets re-engineered over 5 years

Summary

Nikesh Arora, CEO of Palo Alto Networks for nearly 8 years (market cap up from roughly $17B to $238B over that stretch), joins the All-In hosts to walk through what AI capability is actually doing inside a large enterprise, starting with his company's own AI code-scanning tool, Mythos. In 6 weeks of testing, Mythos found what Arora says would have taken 5 to 7 years of human effort to uncover, for a token cost in the low millions, and in "ultra mode" it can chain individual vulnerabilities into full attack paths. He's direct that this cuts both ways: the same capability that lets Palo Alto find its own bugs faster is roughly 3 months, not 6, from being available to attackers in the open, since frontier model weights are now small enough to fit on a USB stick and get distilled elsewhere within days.

Arora spends real time on why raw model capability isn't the whole story for enterprise defense: an early model he cites had a roughly 30% false-positive rate, which makes it useful for finding attack paths but unreliable for triage, since a third of "vulnerabilities" flagged aren't real. He generalizes this to any high-stakes automated decision (insurance claims, self-driving cars) and argues the unglamorous work of driving false positives toward zero, not chasing newer model releases, is where most enterprise AI engineering effort actually has to go. On national security specifically, he reframes the risk: 89% of breaches trace back to stolen credentials, not novel exploits, and hardened critical infrastructure is already well defended. The real exposure sits with small, under-resourced operators, illustrated by the Change Healthcare ransomware breach that froze physician offices nationwide and forced UnitedHealth to issue billions in credits.

The conversation's second half is a software-market thesis. Arora splits enterprise software into three buckets: analytical SaaS is "over," because once a company's data is unified, pointing an LLM at it directly beats paying for a dedicated analytics app; infrastructure software (Databricks, Snowflake, MongoDB, Oracle) is undervalued, since enterprises will need roughly 10x the data storage within 3 years; and system-of-record software (Salesforce-style systems of work) will be re-engineered over the next 5 years as UI disappears and agents interact with backend data directly. He backs this with a live example: Palo Alto killed a SaaS tool only 3 of 20 licensed seats were actually using, wired the underlying data into Slack and Claude, and cut that spend by 90%.

On where the money actually goes, Arora argues profit pools sit in the application layer, not in the base models, pointing to OpenAI's Codex and Anthropic's Claude Code as the parts of those companies actually "running away" with growth. He expects an intermediate layer of application companies to form that arbitrage between models and specific business problems, a layer he says hasn't fully crystallized yet. He also pushes back on the assumption that AI shrinks headcount, saying Palo Alto currently employs more technical staff than ever because AI is forcing transformation across every function at once. He closes on his own M&A logic: after years of buying product companies and running them through Palo Alto's go-to-market engine (most recently a $25B identity-security acquisition closed 3 months prior), he now sees a path where proving superior AI-driven operating margins (90s gross, 40s net) becomes the justification for acquisitions well outside cybersecurity's traditional lane.

Notable Quotes

"In 6 weeks we found vulnerabilities which would have normally taken us 5 to 7 years to find." - Nikesh Arora

"If you're an analytical SaaS company, it's over." - Nikesh Arora

"The entire weights of their most recent model can fit on a USB stick. That's the IP." - Nikesh Arora

"89% of breaches happen because of simple things... credentials get stolen." - Nikesh Arora

"I think we're going to have more people at Palo Alto on the technology side than we've ever had before because I think AI is causing everything to ask for a transformation." - Nikesh Arora